Banking on Fintech? Investors’ POV on Fintech’s Big Reset

Shawn

Hello, this is Shawn Budde of Ensemblex and this is the Ensemblex Exchange Podcast. Today I'm talking with two great fintech VCs, Kwesi Acquay and Noah Breslow about fintech and AI. To introduce you to Kwesi, he's a principal at Redpoint Ventures investing in early and growth stage software, vertical AI and fintech startups. Prior to joining Redpoint, Kwesi was a tech investment banker at JP Morgan where he advised clients across IPO and M&A deals including IBM's $34 billion acquisition of Red Hat. Not bad. He writes about vertical AI innovation and market trends in his blog, Scaling Verticals. Kwesi earned an MBA at Harvard and attended Cornell in gorgeous Ithaca for his undergrad. Welcome, Kwesi.

Kwesi

Glad to be here.

Shawn

And Noah Breslow is an experienced entrepreneur with over 20 years of experience in the fintech and software industries. He joined Bain Capital Ventures in 2021 as an operating partner. Prior to that, Noah helped build out OnDeck, a pioneering online SMB lender, as the chairman and CEO. He previously ran marketing and product for Tacit Networks, helping lead the company to a successful acquisition. A graduate of MIT, Noah is still an advisor to MIT's Delta V Educational Accelerator and is also a mentor in the NIC FinTech Innovation Lab. And welcome, Noah.

Noah

Thanks, Shawn. Great to be here.

Shawn

So I wanted to start by just talking about how you found your way into the VC industry. Kwesi, you went relatively direct, Noah a little less so. So Kwesi, why don't we start with you? Like, what is it that drew you to being a VC and joining that industry?

Kwesi

Yeah, I'm really glad to kick things off here. I think one thing just from hearing everyone's stories, there's no one standard way to go into venture. And I think everyone's story is a little unique. And so for mine, I started my career at JPMorgan in tech investment banking. And so really a lot of software and fintech. As I've noticed across my career, timing is a little bit of everything. So we're talking mid, late 2010s, fundamental shift to the cloud, emergence of fintech. And so really enjoy the opportunity to work with a lot of founder-backed businesses. And that was really that initial spark that got me really interested in understanding, hey, what's happening with founders that they're able to create this value? There was a little bit of a different ethos around vision. And I also got a front row seat of just how large incumbents were trying to figure this stuff out. How does IBM respond? How does JP Morgan even internally experiment? There were things that JP Morgan said that in 2014, hey, we would never put you know, this specific company in our tech stack. And then, you know, you go talk to them, you know, a couple of years ago and it's going to be like 90% of their workloads are going to be in that system. And so the inflection points were always what was interesting to me. And I think the partnership with founders was something that I was excited about. And so spent six years in banking, went to business school was a great reset for me. And, you know, I was particularly intrigued about investing at the early growth stage. So call it series B. And so, where I got excited about that stage particularly was it's a unique inflection point where, you know, I think that's the difference where you're trying to figure out, is this a great product or is this a great product that can eventually become a great enduring business? And so certainly you try before you buy in a sense. And so internet multiple venture funds certainly was peak market activity. So there was a lot to look at and a lot to be excited about. And then I ended up at Red Point in 2022.

Shawn

Noah, I remember when you first joined Bain, you didn't know if it was going to stick. Like you were kind of, I think you were halfway in on it. So, you know, now you're fully in. What did you see that maybe you didn't expect to see that got you to the other side?

Noah

Yeah, no, I think I went from being like an accidental tourist in the world of venture capital to obviously something much more permanent and long lasting. Yeah, I had a 20 year operating career and operating venture backed company. So I had worked with venture capitalists most of my career. I had raised hundreds of millions of dollars from venture capitalists, especially for OnDeck, as you mentioned. And I felt like after a 14 year run at OnDeck, after we sold the business, I felt like I had thought about one problem very deeply for a long time. And what I was sort of hungering to do was go broad. And I think what's exciting about being on the other side is I think within my first six months at BCV, I had looked at like 150 different companies, got much broader exposure to fintech and software and commerce and some of the other areas that we focus on here. And so when it came time to make it a more long lasting relationship, I really enjoyed kind of being on the other side. And I still am. I think it's a very, very intellectually stimulating place to be. And we'll talk about where we are in the cycle on different technologies. But there's always something new. And I think that's what's cool about being on this side.

Shawn

Yeah, so let's talk about that, where we are in the cycle then. So what are your thoughts on fintech in the US specifically at this stage? Is all the juice gone? Is there still more to be had?

Noah

We're still long fintech at BCV. It's still a very active domain for us. Probably about 30% of our investments are in or in the fintech space. But you have to choose your shots because the white space is being reduced. I think we have a wave of, you know, the fintech 1.0 company, Shawn, that you and I know very, very well. We've got 2.0 companies kind of like the Affirms of the world, that are public and doing pretty well. And so I think if you're starting a company today and fintech, you really have to decide like where are some of the uncharted areas or untrod trails, so to speak, that you can really make an impact because, you know, fintech, it's funny that with the tooling that's available to today's fintech entrepreneurs, there aren't that many barriers to entry, but there are big barriers to scale. And I think very few companies can kind of get to that escape velocity and public scale. And so that's what we're kind of looking for. So we still see opportunities in insurance. We still love payments, excited about the possibility of stable coins there. We like the office of the CFO. We think there's a lot of tech transformation happening in that tech stack. There's the perennial needs for identity and fraud defense and better data. And we're seeing some things in like tech enabled services too, that when you combine AI with them, you can get some pretty interesting outcomes. But again, got to pick your shots. And certainly the public market valuations aren't helping the case, especially if you're at the growth stage right now.

Shawn

Yeah, so Kwesi, what are your thoughts on that?

Kwesi

Yeah, I think even just thinking about the backdrop here is important context, right? So I really think 2023 was somewhat of a reset year for fintech. So if we take pitch book data, you know, 35 billion of venture funding, you know, went into fintech, that's down 44% year on year, you know, probably a more apt comparison given there was a peak in 2021. If we even looked at pre-pandemic levels, say 2019, it's still down 20%. And so, what's driving this? So there's a couple of things, right? So certainly a challenging macro environment. High growth is hard to achieve, as Noah mentioned, in terms of scale. Certainly customer acquisition costs, loss rates have risen. And so, there's a set of startups that may have raised, for example, in 2021, 2022 at peak valuations that they're still trying to figure out the right way to go get that scale to grow into those valuations. And so there's certainly public market activity that's influencing private market activity. I would say that if we just looked from a mass perspective of just like where growth investors are going, a lot of it has been somewhat of a perspective of flight to quality. I think 70 plus percent has been more B2B versus B2C. But as Noah mentioned, there is still a lot to be excited about, and so there's certainly a lot at the earlier stage of investing where there's a little bit of this reset even from a builder perspective. And so, people are looking at different ways to think about payments, CFO tools, right? Audit workflows, insurance, and certainly AI sometimes is a part of it, but it's not the full story. I think there's been a really good reset where, you know, folks have been really thoughtful about, hey, if I'm gonna go build a fintech business, it needs to be pretty unique in how you wanna build and what end markets you choose because it's certainly a little bit more of a complicated task than the typical run-of-the-mill software startup that scales.

Shawn

Yeah, I think people underestimate the complexity of fintech. I think we have a lot of folks who came in from the outside and this is all just a tech problem and everybody who's doing it has been doing it wrong and therefore I'm going to fix it. I want to go back to something you said before though, Kwesi. You said, It's a great product that grows into a great company or something like that, right? So can you talk about that? What does that mean?

What's the difference between the product and the company in terms of the way you think about it?

Kwesi

Yeah. So, you know, it's certainly see that at the series B perspective, right? It's you've gotten some product market fit. Maybe you're scaling, maybe you tripled, you know, your run rate revenue or your ARR. And so something's clearly working here. And so, you know, you're the product is certainly resonating with a core audience. Let's say, you know, business is sitting at, 5 million of ARR. It's had a really attractive growth. Now the question becomes, how do you go get, that growth of how do you turn five into a hundred, 200 of what you think the, you know, the average, you know, IPO-able candidates are. And some of that's a lot of different factors. So sometimes that's, Hey, do we think you could build multiple products on top? Some of that is how do we continue to grow new customer segments or grow within the existing segments? How do we also be more reliable? Right. And so that's a little bit of the sustainable growth of how do you have really attractive retention over multiple years, all the while just making sure you have an understanding of where to put your investments because if things are starting to work for you, others will follow and your competition, whether it's from incumbents or even new startups, will evolve. And so I think particularly in the context of fintech and I know we'll get to lending in a second, those vectors of expansion revenue, those expected differentiators around against competitors, I do think is quite unique. And so, you know, when you start to see a product, you know, that that's really resonating in the market, you want to say, OK, like, how does this resonate even mass broadly to be this great business?

Shawn

Yeah Noah, did you want to add to that?

Noah

I think a lot about this issue. When I was a builder, I thought about it, and as an investor now, we think about it too. It's like, you've got features, you've got products, you've got companies. And over time, all products become features, and all companies kind of degrade down to products, right? So the market expectations like keep ramping up over time, right? You think about your iPhone today and all of the separate products that are now collapsed, right? There used to be GPS devices and there used to be, you know, cameras, right? And those things are now features of an iPhone. I think that's true in a lot of tech-based businesses. So, you know, I think what we're looking for in some ways is how sustainable is the distribution moat that a company has, right? Can we lock in on a certain customer and can we reach that customer in a cost -effective way that then allows us to layer in on our core product that might be giving us that initial revenue traction that Kwesi was talking about. How do I layer in other products and services that kind of interlock and create a whole that's greater than the sum of its parts? And so I look at a company, like a great example is in my mind is like a Carta, for example, you see what Mercury is doing in SMB banking, but like in Carta's case, started out as CapTable management. Well, who are you serving? You're serving the startup founder. It's a great entry point because people adopt Carta right when they're setting up their businesses. But then you radiate out from there and you start doing 409 A valuations. You might think about selling the data that's on your platform to investors. You think about fund administration for the venture investors who are holding shares on the Carta platform. You start rating out now not just serving one customer, but actually two sets of customers, the VCs as well as the entrepreneurs. I think examples like that, you see it with Ramp to some degree. You see it with Rippling. You know, some of the real fast moving companies that have grown like crazy. They just keep figuring out how to add components to that initial value prop to then create this kind of unassailable mode over time and customer lock-in.

Shawn

So, Kwesi, you said we'd get to lending in a second. I guess that second has come. Where do you two stand on lending businesses? It seems that everybody wants to sell or I guess invest in companies that are supporting lending businesses, but nobody seems to want to invest in the lending businesses themselves these days. So where do you guys stand on that?

Kwesi

Happy to jump in here. You know, I think there's certainly a distinction here of if you're a pure play lender and then there's also, you know, there's, there's a lot of really interesting software businesses that also embed lending as part of the, the sell here. And so, maybe just talking specifically about just some of the challenges of, of the pure play lending, which I do think it's, it's, it's possible to make a great business out of this, but it's certainly very difficult, right? And so we talked earlier about some of the macro factors, right? You have to deal with regulation, right? Rates driving up cost of capital, right? You have to think about policy. When you start thinking about like relative to, you know, use it as an anchor, kind of your classic software startup, how you diversify the business is a lot more complex, right? And so there's an element around, how do you go get that expansion revenue? And so that sometimes means, adding different lending products of different asset classifications. There's an element that I do think takes a lot more discernment; it's product market fit isn't necessarily automatic, and so that's certainly another element. And the last thing I'd say is just as you add new customer segmentations, there is this delicate balance of speed and quality. So when you think about, hey, can we move fast, attract new customer segmentations so we can drive growth? But then that has to balance with your underwriting model and say, hey, like for these newer cohorts, do we understand it well enough to also maintain our underwriting advantage and then hopefully have somewhat of a flywheel to go. And so I think there's always kind of this balance around how do you get high credit quality in your loan portfolio because it's, you know, having, you know, three X-ing customers in software is very different necessarily than three X-ing as a lending business.

Shawn

Yes. It's easy to grow as a lending business. It's the getting paid on the back end that is the challenge.

Noah

Exactly. Yeah, no, I mean, just to build on what Kwesi was saying. Look, lending businesses are tricky from a VC standpoint, and we don't do a ton of them here at BCV, ironically, because I have a background in it, and one of my partners, Matt Harris, obviously was on the board with us at OnDeck. But the failure modes are tricky of lending businesses, right? So that's one issue is that they don't sort of gently die. They have kind of a violent death when things don't go well. And then there is this weird trade off, like you sell your product, you don't know what your margins are for the next, you know, X number of months or years until the loan pays back. And then you have this strange trade off between, you know, revenue and profit quality and growth. And so, VCs are looking to invest in these breakout unicorn power law type businesses that you could 10 X your growth in a year. And you're a happy camper. As you said, Shawn, if you 10 X growth a lender in a year, If you could short the private stock of that company, you probably would. So that's a bit of a tricky dynamic. We are excited about, we do like the tools that are sold to lenders, to your point. That's a little bit more of like the arms dealer place in the market. And then you can sell into banks, you can sell into established lenders, you can sell into startups. And I think that's intriguing. We do like the embedded lending model as well. So we have a company in our portfolio, WiseTact, it's an embedded consumer lender and they partner with a lot of different platforms, a lot of banks to enable a more frictionless lending experience as a feature of say a vertical software platform that's serving contractors. And so that model is fascinating because you get a little bit more of an edge, right? You get the data from the platform about say how a contractor is billing their clients, their longevity on the platform, do they reliably pay their bills? And you get an edge on customer acquisition because the platform's already got that contractor using it for, let's say, tracking their jobs and invoicing customers. So the data plus the lower customer acquisition cost, we think leads to some sort of structural advantage that sort of a lender just acquiring customers in the open market just doesn't enjoy.

Shawn

Yeah, that's something we've always looked at; what is the information advantage, in essence, you have that allows you to underwrite better? I feel like embedded lending does offer that. It also feels, and I'd love your thoughts on this, I've always said you want the multiple of a software business, but you want the earnings of a lending business. Do you think embedded lending helps you thread that needle, or do you just get valued then like a lending business?

Noah

I think it depends a little bit on your risk model, right? Do you hold the assets? Do you have a way for someone to take that credit risk from you? But I think you'll be valued at the high end or maybe even a bit higher than the highest kind of standalone lender. That's my feeling is you probably won't get the pure tech in multiple because you are still a lending business. There are cyclical considerations and macroeconomic forces that you have to contend with, but that structural advantage on data and distribution should lead to a more predictable earning stream should lead to a higher multiple and more frictionless growth without sacrificing credit quality. So I think those things are really positive, but the idea that an embedded lender will trade like a snowflake, I think is nuts, and you have to watch out for that.

Kwesi

I would say, some of these software players that have a lending component, there's some level of like order of operations, right? So for example, if you think about, you know, a Toast or Softrify, it's like they technically have embedded payments, embedded lending, right, as well. You're really thinking about like the software leading. And so I do think as you start to think about building it, order of operations, I do think is going to matter a lot because when you start thinking about maybe you're, you're starting with software, you have a view of where your advantages are. You have a view of kind of the stickiness. And then lending can be, it's, it's a heavy lift, but maybe you're at scale, you have more resources and you can think of very measured approach to what ICP is going to matter. Obviously from a valuation discussion, it's, and I guess putting my former banker hat on, it was like, there were so many businesses where it was like, you're, you're arguing for the software multiple, but they want to give you the lower payments multiple. And, you know, there's definitely an exercise in, kind of digging deep into the composition of the business. I think fundamentally it just comes down to just what's truly recurring. And I think there was a conversation around payments about this as we went from 2021 to now is that going off of like, you know, interchange, and run rate, like GMB, it's like, that's not necessarily like, that's not ARR. And so it's not contractual. And so there's some layer where that's typically the math that people are trying to figure out is just like, how recurring is this business? And so if you have more proof points to figure out, okay, here's, you know, how we think about repeatable business and you can sometimes lead with software. I do think that gives investors confidence and sometimes I can end up in a, in a multiple, uplift.

Shawn

All right, so let's go back 10 years. Later, I'm going to ask you to go forward 10 years, but let's go back 10 years and think about where we were on fintech, at the time. What was driving the hype at that point? We've seen things come down, right? Lending clubs down 90 % Affirm, 80 % upstart, 90%, SoFi 70 -ish. So we've crashed. We're on the other side of the hype cycle. What was driving the hype cycle back then? What got people so excited?

Noah

Yeah, you know ZIRP would be exhibit A in this detective case here. Zero interest rates for that long a period of time creates an environment with heavy demand for capital. It creates an environment where customers are going to default less and ultimately makes equities look very, very attractive relative to their fixed income counterparts, and so all of those things, I think, conspired, plus a little bit of consensual hallucination, right? And Kwesi just nailed it. Like people started to say that all revenue is created equal, but the sort of fundamental laws of valuation physics still apply. And like the highest form of revenue is that recurring software, like lock-in type revenue. But you start degrading from that. You know, you start degrading both in the level of recurrence and you start degrading in the level of margin quality and in the level of ability to sustain that revenue through different cycles. So, you know, I think people sort of forgot that they're like all revenues created equal, lending companies should trade on revenue multiples, which is bananas. And then the ZIRP factor gave everyone kind of these crazy growth rates and like, you know, having lived it, having taken a lending company public at a tech valuation and watched it start to trade as a financial services company, which is what it should have done all the time. We had a call with an analyst and we were talking about lowering our growth forecast for a variety of reasons. He said, look, you guys are still unprofitable and all my value for this company is in the terminal value. And so if your growth comes down a little bit, and let's say your profitability timeline goes out a little bit, like it has a massive impact on your valuation. So you look at a company like Upstart that was trading at like, what if it was $400 a share in the fall of 2021? And then I think, I don't know where they are today, but you know, tens of dollars, like, you know, they've come down a ton. A lot of it is that when the growth rate starts coming down, the credit quality starts to suffer that future quality of earnings outlook changes so dramatically, and if you're not profitable, it's even worse because there's no current period earnings to kind of support a floor on the valuation. Those are really, really tough resets and you've seen all these companies now go through it.

Shawn 

Like part of what was interesting to me at the time is everybody loved platforms, right? We don't have all the contingent risk on the backend, blah, blah, blah. You also don't have the revenue. Pretty much every month, right, you started $0 of revenue because you're earning mostly off of commissions, which feels to me, you know, I understand the charm of not having the credit risk, but you've got a very different set of risks. And to your point, Kwesi, that's not recurring revenue at that point, right? That's just this month's revenue.

Kwesi

Yep, yep, and doubling down on Noah's point, around just the impact of low interest rates, what you effectively had was really high mark and market demand was basically like fixed for a very long time. And so if you think about a lot of the names we were just mentioning in the prior period, a lot of those were 2021 IPOs. And so if we kind of went back and we looked at the S-1s of those businesses. You know, if we're anchoring on, okay, everything's going up, there's clear demand. and then you couple that with just really a very common theme. So these kind of hidden large TAM buckets and this, this notion of it's still not optimized for credit. So sometimes it's searching for the hidden prime candidate. Sometimes it was a little bit more on SMB. The other part I think was, I think there was a real belief around just these data plays and data flywheels that could thread the needle with underwriting. So you could drive at the same time, more approvals, fewer defaults. But part of that evidence was somewhat commingled with the zero interest rate environment. And so really, I think you mentioned platforms before. I think one of the challenges as an investor to really think about is, I think we love to talk about platforms. I do think it gets a little tricky when you start to think about data platforms and data moats. I do think that those graphics of just the circular benefit and it doesn't always hold that way, and so I do think that, you know, there's some precision that, you know, like what's actually driving performance here or, you know, if you're talking about alternative data, sometimes only a couple of the metrics actually drive really the insight versus more data equals, you know, significantly more value. And especially as you start to get different customer cohorts, you're not necessarily getting the transferability across. And so I do think there was some level where even the data play was a little overstated in the market when it was really, you know, a lot of the macro factors drove a huge part of that boost. And so that wasn't, in my view, it wasn't as well accounted for. But certainly, you know, I think at the time, probably like 20% of VC dollars were even going into fintech. So it was certainly a view that, you know, this was an area that was ripe for disruption versus traditional banks and all that. It's just hard because it's, I think, part of the element around you know, lending and product market fit is like you have to kind of go through a business cycle to actually feel like you have this viewpoint around if this is actually going to work over the long run.

Shawn

Yeah, I remember I was at an event where they asked a bunch of fintech entrepreneurs when the next recession would come. And when we got to 10 years, I think we still had less than 10% of the room had their hands in the air. It was like there was just this belief that it was just that we lived in a different world that was going to avoid, I don't know, the cycles that have occurred probably for the last several millennium. So why don't we talk about what is disrupting things now, which is AI. So Noah, I believe you and I over Moules Frites, ten years ago in San Francisco first talked about Machine Learning AI as it applied to underwriting. I think you took some of that on with OnDeck, and I'd love to just kind of hear about, I guess, version 1.0 of AI and how you guys use that at OnDeck and how you see that having evolved into the new version that has become much more well-known in the public eye, at least.

Noah

Yeah, no, that's right. I mean, I think, you know, people talk about AI today or generative AI. And back then, I think we were all talking about machine learning. You know, so there's supervised and unsupervised, you know, techniques and both have relevance in these types of financial services businesses. So, we heavily use supervised learning to build our credit models. I mean, you know, the OnDeck score essentially was a labeled data set with inputs of all of our loan applications, and we were predicting the probability of default for our customers. And that score steadily improved over the course of a decade. Took about a billion dollars of loans for it to start being really predictive in a differentiated way. By 10 billion of loans and kind of a history of hundreds of millions of bad loans behind that, we were able to really get enough data points to build something we felt that was fairly defensible, fairly proprietary, and that score still supports the OnDeck business today, which is securitizing in the market. So that was sort of the supervised piece. And then unsupervised was like cooler, but weirdly it was the supervised piece that drove most of our economic returns. But the unsupervised stuff we used more for things like feature discovery. So we're collecting all this data and we're trying to figure out what features might be interesting to include in our supervised models. And so we would use clustering techniques and other things to try to find features that would either give us sort of an orthogonal predictive value to the stuff we were already using, or maybe try to find features that were sort of doing the opposite of that actually, like were quite highly correlated and we could simplify our models by getting rid of some of them. So, we saw a role for both, but you know, LLMs are a different animal altogether. They're both unsupervised and probabilistic, right? So they don't necessarily give you the same output twice to the same question, which is kind of fascinating. And so I think what's exciting about it is, like, I don't know if I would trust an LLM to build the OnDeck credit model today. Not yet. Maybe at some point, or maybe it could instruct other tools to build the credit model, but I'm not sure the LLM technology is really that amenable to it. However, there were so many other process-intensive tasks in lending and financial services more generally where LLMs can apply and anything that's kind of text in, text out, eventually it'll be voice in, voice out or video in, video out. That's what's exciting I think now is that we can apply those to a lot of the operational activities in financial services and it's a very operationally intensive industry.

Kwesi

To add to that, you know, I think if we were just talking about really the, the technical spikes around just generative AI, I'd say there's a couple of planes. So really just the ability to handle and structure data. And so where there isn't necessarily a clear delineation and relationship, and that can help you figure out, Hey, so there are some new relationships that, we, we haven't been able to figure out with a little bit more of a predictive model. I'd say secondly, there is an element around more nuanced semantic understanding. And so, that certainly helps from a user experience perspective of, I can type in just, hey, this is what I'm looking for, just as if I was just writing an email and it's going to understand things. And then also in the sense of, you know, with some infrastructure tooling, you can even get it bespoke to your industry or bespoke to a specific customer of this is how we think about things in the insurance industry. This is how we think about things as X firm. And so that element I do think allows for this step function. And obviously, you know, we talk about like chat GPT and that's exactly where you get to see that explosion is that there's a lot of elements around transformer models in the tooling layer, but it's mirrored with this ease of use at the user level. You know, certainly there, there's an element around just, you don't need to write the full rules of exactly the steps. And so that opens up a lot of possibilities. You can certainly provide directional instruction around prompting. The other thing I'd say that's been more prevalent over the last couple of months is also just multimodal capabilities. So I know a lot of times what we're talking about, LLM’s, it's really about text, but really over the last several months, we've seen just incorporation of video. We've seen a lot of incorporation of other data modalities. And a lot of times it's the mix between the two, right? If you think about, you know, syncing video and audio together and having just an understanding of that. And we've seen that in a number of different industries. And so there's a lot to really like about what's happening in this space and certainly driving a lot of the hype and... you know, I think it's, I think fundamentally what gets people the most excited is, is we possibly could be, you know, in this world where, you know, generative AI could potentially be the greatest force multiplier we've seen, you know, from productivity perspective, the fact that it can really address things at the enterprise level, SMB side, and a plethora of different industries.

Shawn

Yeah, so I mean, you guys see, Noah, you already said it, right? You see a lot of pitches where I assume AI is written in at least every deck, at least one place these days. There is going to use AI in some way. So where do you think it really is going to add the most value in fintech type businesses?

Noah

Yeah, I think as I mentioned, the operational side of fintech, I think is, it's funny in the fullness of time, we'll see if that's where it adds the most value, but I think that's where it's gonna add the most value in the near term. Cause in some ways it's the safest place for humans to allow AI into the tent. You know, it's tough to go into a bank and say, we're gonna get rid of all your judgmental underwriters that have been approving loans for you for the last 10 years, right? There are regulators wired into that credit process.

Shawn

Right.

Noah

There's obviously a huge balance sheet at play with the loans that the bank originates. And so it's hard for them to take on that risk. But what's unassailable is we will take that underwriter and we will make them five to 10 times more productive. And when we looked at this, you know, at OnDeck, it was like 60 to 70 percent of our underwriters time was spent assembling information and teeing up data that was in a variety of different formats to ultimately make that judgmental, you know credit decision where we use judgment. We used automation for the smaller loans, but for the larger ones we use both and so I think this productivity sale is a lot easier than a replacement sale and so that's that's where people are gonna focus. And I think it does come back to what Kwesi was talking about which to me is one of the most exciting things about this whole technology is Gen AI makes software less brittle and the moment software is written historically it starts to become dated and one of our portfolio companies, very simple example, like they have a business process where they need to ingest an employee census file, right? A list of all the employees and how much salary they make and what their start dates were and their social security numbers and all that good stuff. Five years ago, you would have to say, okay, there's 200 payroll companies out there of interest and we need to build parsers for every single possible kind of employee census file that we might see. And that would be a very labor intensive exercise. But today you can say, with a prompt, here is an employee census file. It typically contains these fields, you know, do your best. And if you're not confident, let me know. And all of a sudden, 95 to 99% of those census files are handled for you. And you didn't have to write a hundred parsers and the software doesn't get out of date when a file format changes. And I think that it's one very, very specific example, but like, if you sort of multiply that across all the places in software where different systems are talking to each other and fintech's very, very heavy with this.

You can imagine a much more plastic landscape for software than we've ever had before. And that I think should lower costs for people building in fintech and it should make the systems a lot more resilient.

Shawn

Yeah, I mean, we launched a company called Gestalt, which is organizing lending data. And their secret sauce is that they've built that AI engine to do 90% of the mapping in a, like say a semi-automated way, right? It says this one, whatever, account number is account number. I'm pretty confident on that one. You know, with this other variable, maybe not so much. I need someone to look at it. Kwesi?

Kwesi

Yeah, I think Noah's comment on just really the unlocks for software, I think is one of the best ways to really think about the potential even on the fintech side. So if we take financial software, for example, there have been VCs for years about just we're going to go build this software platform. It's going to understand its customers. And then there's this piece on top of it where we're going to get the benefit of the data. And we've been waiting a while for that. And I do think there's a credible argument that generative AI can make that piece. And I think there's going to be, it's going to be a multiple step process. I think where we're at right now is a little bit in this like copilot era where there's some querying and response. And so it's a little bit more of, you know, as Noah was referring to, just kind of this like, almost like agent assists, right? So it's human reviewer, right? If it's, you know, like on the insurance side, you know, example would be like a six-fold, right? And so based on your preferences and your policies and how you think about risk, let's surface up all of the different dials. I'm not going to make the decision for you but I am going to surface all of the key criteria upfront. So you're saving incredible amount of time. At the same time, you're allowing the human to provide a review, which certainly you're learning from and of the actions. What the next step function is, and I do think that back half of the year, more people are going to be talking about really this concept around AI agents. And so it's not just about, Hey, this is a Q&A it's, Hey, I'm basically going to give you instructions. Like, you know, this is kind of your mandate. This is your expertise. And I need you to be able to run this full process, whether it's, you know, KYB, whether it's, like CFO tools, Hey, write me, you know, a scenario plan, right, if you're talking about like CFO tools, and so I think that's really the next step function. I think what's incredibly interesting about this time period is that it feels like every four or five weeks, there's a step function and what state of the art even means. And so I do think that's going to be an interesting parallel.

Now, it is incredibly difficult to build AI agents for a ton of different reasons of, you know, how do you avoid recursive loops? How do you think about making sure that, you know, you're able to still connect to the right systems online into core systems. So there's a lot that needs to be built there, but there is certainly a clear line of sight around just this ability to help in the near term productivity. And then a little bit more in the medium to potentially long-term thinking about there could be whole processes that get automated and then really your job as the person is almost like you're the supervisor. And so it's really only elevating when there's a little bit of a trickiness. And so, you know, I think that is something where we're really headed to, but there's a ton of use cases that people are experimenting with, I think in the finance sector, and so I think there's a lot to like about it.

And then, you know, the last thing I'd say is, you know, from an underwriting perspective, you know, it's probably unlikely that there's going to be a replacement of the core system, there's just too much complexity, too much regulation. I think that there's a lot of risk there, but what it can do is supplement. And so, you know, like companies like Slope are basically using transformer models where they're able to just make better sense of really complex and irregular bank transaction data. And then that feeds into their core model. And so we will see more things like that where, Hey, how do we make sense of this data that we've always struggled with? Hopefully there's a net effect of your core models get better because there's a little bit more data enrichment, data insight that, LLMs and others are able to build on top of.

Shawn

I think you're right. The, the, we're in that co-pilot era. It can make us more efficient. I actually, on a recent business trip, I finally, after five years lost my Kindle. and I didn't want to go pay full fare for the Kindle. So I wanted to figure out when prime day was coming. I asked my search engine that came back and told me Tuesday, the 23rd of July. It then told me like what day of the year that was and there's a hundred and some odd days left. And then it ended by concluding that Prime Day was most likely to be on July 9th and 10th. So it gave me conflicting answers and a lot of irrelevant information. I think we're a ways off.

Again, I think if you have a human underwriter sitting there looking at that, they can kind of say, OK, I've got two answers. I can figure out which one's the right one. If you leave your underwriting to the machine, you would be very sad.

So what I wonder about this is, you know, the pitch 10 years ago, I think for every finance company said they were democratizing credit, right? Democratizing was the big word I heard a lot. Is AI going to democratize things or is this just going to be a tool that helps big companies who can afford to invest in it and really understand how to leverage it? Is this just going to help them get bigger and push out the smaller companies?

Noah

There are going to be some big companies who win in this for sure. You look at Nvidia and how much the market cap of that company has appreciated, and that's sort of a harbinger, I think, of the entire AI market in a lot of ways. They are the sole provider, essentially, at this point, for the raw computing, which is one of the key ingredients here needed. Soon we're going to get to an energy shortage, probably, but for now, the computing is front and center. So you look at Nvidia, you look at Microsoft, you look at Amazon making a huge investment, I think in Anthropic, you've got Google. So I mean, these players have been winning in tech, right? There's a reason they're called the Magic Six or Magic Seven, and they're going to keep winning. But I do think Meta's move of open sourcing models and investing heavily behind that, and also a little bit of like the arms race between these big companies is rapidly driving down the cost to access these core GenAI Models. So I think between the open source models that are getting higher and higher quality kind of every month, and then this sort of arms race that's going on, a lot of midsize and smaller companies could just build on top of those fundamental layers, much like, you know, small and midsize companies built on top of Intel back in the day when, you know, it made the PC a more affordable thing. And I think we're in one of those waves. And, you know, it's funny, we are a venture firm that's attached to a large private equity firm. And a lot of our private equity portfolio companies now, these are not super tech forward organizations, right? A lot of times their tech is going from the past to the present. But that being said, they're able to deploy AI in their businesses in a couple quarters. And so the time to utilize this technology is a lot lower than other tech trends that we've seen. And so, you know, you can activate those use cases pretty quickly. A lot of these companies have heavy body counts doing very manual tasks. And so it's not that hard to see the uplift. 

So I guess I'll give you kind of a democratizing answer. I sort of think, yes, the big guys are gonna do quite well, but I don't think it's like there's only gonna be table scraps left over for smaller firms.

Kwesi

I think another way you can think about is actually splitting it by the stack. So if we say AI models and compute, then yes, I mean, just the amount of capital you need to really make that go in a recurring basis to continue your lead. You know, those are billions and billions of dollars, and that's why you've seen basically every AI model have some massive big tech corporate engine, whether it's a special relationship, whether it's actual explicit funding. And so...that I do think is going to be a large company game for sure.

When we're talking about applications, I don't necessarily think it's going to be true. And so that's kind of the difference of, you know, if you're a vertical software company, that's super large and you know, you're adding generative AI as a new product versus maybe someone that's, you know, GenAI native and that competition going on. And the reason why I say it's not necessarily true of just like the bigger company with, you know, a hundred customers, is going to beat out someone that's GenAI native is one is when you think about data, like more data doesn't always necessarily mean better model performance. And I think there's a lot more research that's supporting that fact now. And so, and even, you know, on the perspective of just data quality, right. And so just because you've had, you know, 2012 to 2023 data on customers, the effort to get the data in the right position to go build from a model perspective, I think that's, you know, a harder feat than most people understand.

The other part I would say is just, you know, as Noah mentioned, there's a lot more wave in open source. And so there's a lot more choice for folks to specialize. And so there are going to be some use cases, right? Like, you know, Mistral is using a mixture of experts, right? So small models getting a lot of great performance and you're always balancing performance and costs or, you know, other elements that you really need to care about for your specialized use case. And so even at the infrastructure layer, a lot of startups are getting choice to figure out what's best for them to optimize from a cost perspective. Training costs are going down, inference costs are going down, and so there's a lot to like where more folks can kind of get in the fold. And I would also say there's some element in just in tandem thinking about just how fast things are moving. It favors folks that are exceptional at product velocity, right? We just talked about AI agents, and so if you can create great software and then you just have highly performing agents, which is an incredibly technical feat to do that potentially could be a leapfrog effect over maybe an incumbent solution.

Yeah, I would say at the time, you know, as we said today, I mean, a lot of the GenAI solutions, whether it's incumbents or even startups, a lot of it's still in the pilot phase. So not everything is locked and loaded in terms of signing contracts of, hey, I'm going to use this for five years. And so there's plenty of opportunity for folks to come in into play. Now, you know, certainly there's a go to market advantage you have to overcome here and prove your enterprise great as a product and really push through, cause you're not going to be able to plug into core systems unless you're really proving very core things around security. And so there's something that's really important there, to consider as you think about products, but there's plenty of opportunity to really have a high product velocity products. I do think there's plenty of opportunity to really think about how can I, you know, build purpose built software or solutions, that adopt not just the latest technology for the sake of it, but it's actually complementing some of the needs of, you know, industry players. So certainly if we're talking like the core core of fundamental chips, compute AI models, big company game, I think it remains to be seen on the AI application layer.

Shawn

Yeah. All right. Well, thank you. Let me just close with this question. I promised I would ask you to go into the future. So I'm going to ask you to go into the future 10 years and then look back at today. Are we in the middle of a consensual hallucination or how are we going to be looking back on today in 10 years? What are we going to think about AI and where we're going to be at?

Noah

So I don't think we're hallucinating. I guess I'll start with that. I think we're gonna look back on today very similar to how maybe we looked back in, I don't know, 2005, 2006 on the web browsers sort of coming of age in the mid-90s, right? It was a major, major technology step function moment. But just like we probably couldn't have picked, you know, or foreseen all the services that would be available in 2005, 2006 from that 1995 kind of vantage point, I think we will struggle to see all the potential, you know, that that's going to come. So I think what we're doing today will look like child's play. It'll look like dial up modems, you know, and America Online and stuff like that. Right. And, those LLMs were cute. Right. Was as a robot, like lurks around our house and does our dishes and, you know, converses with us in fluent English. But, I think it will be recognized as an awakening and I don't think this is sort of a flash in the pan or one of these hype cycles where the technology is not in use in a couple years. So yeah, I think profound change, I think the classic Bill Gates quote is people overestimate the amount of change that happens in two or three years, but underestimate what can happen in 10. And I think we will be profoundly surprised by how much change happens over the next 10 years.

Kwesi

Yeah, I certainly agree on that front. You know, if we do retrospective on, you know, we go, you know, 10 years forward and just think about it. You know, I do think we are going to identify durable use cases. I do think enterprises are going to find a way to make this work. And I think, you know, some of that element is because there are a lot of different ways this can win and you're not really betting on one specific archetype of product. There are even some things just maybe some enterprises figure out a way to go build certain use cases themselves. And there's, it's not an all or nothing proposition. It's you can incrementally work your way towards ROI and just the functionality and capabilities of AI models, AI agents. I do think that's going to proliferate significantly. The other piece of it is I also do think what's going to be really helpful, I would say even over the next two years is that there's going to be a better playbook for vertical AI apps.

And, you know, I've spent a bunch of time just reflecting on this where I think sometimes it gets a little too analogized to vertical software. And it's just so different in terms of how you have to think about building product. And so I do think for founders, just there'll be some lessons learned about, how to really think about winning enterprise customers, how to think about product development, you know, GPT five launches and what do you do about it? How do you test if you need to use it or not use it?

Those are all playbooks that are getting written right now and there's no clear answer. So I do think when we look back, we're going to clearly have the durable use cases. I think we're going to have really great playbooks that I do think founders, early stage founders can leverage and really avoid some common traps. And, you know, hopefully by that time we have a number of generative AI IPOs. And so we'll also be able to see how they perform as public companies. What does run rate margin look like? How does defensibility look like when you got the resources of a public company as well?

Shawn

Yeah, I think one of the disservices of AI and machine learning is the branding of machine learning and AI. You still need the human at the keyboard. And I think what I've seen just with my family, with my coworkers is some folks have really embraced it and figured out how to use the tool. And some folks don't really even know what we’re talking about. So a machine can't do everything for you, but it can certainly make you a lot more efficient. I think that's where we're at.

Well, I'd like to thank you for joining today. You can follow Ensemblex, Kwesi Acquay, that's K -W -E -S -I -A -C -Q -U -A -Y, and Noah Breslow, that's pretty much spelled the way you'd expect, on LinkedIn. You can also visit us at Ensemblex.com, and you can find The Ensemblex Exchange Podcast on all major platforms. Thank you for joining me today, Kwesi and Noah, and thank you all for listening.