Innovating Spend Management through AI with Pedro Franceschi from Brex

0:05Hi, listeners, and welcome to KnowPriors. Today, we're talking to Pedro Franceschi, the co-founder and CEO of Brex. Since its inception in 2017, Brex has been building fintech tools that make employee and employer lives easier when tracking and paying expenses and beyond. So it makes sense that they're now implementing AI to make this process always annoying, even more automated for their enterprise clients. We'll talk about that and the business overall. Welcome, Pedro. Glad to be here. Thanks for having me. I guess, yeah. So just to set context, I think the vast majority of our listeners know, but for any new customers or potential ones, can you tell us what Brex does and what scale it's at today?

0:40Sure. Brex is a corporate card and spend management platform, and we help companies make better financial decisions and push every dollar of spend further. We serve companies from all sizes, from two founders out of YC all the way to large public companies like DoorDash's of the world, Coinbase's of the world, and everything in between. We have tens of thousands of businesses as customers. One in every three startups uses Brex today. 130 public companies are also in Brex. So we have this sort of wide range of customers, which means we're pretty proud to be able to scale with our customers from inception all the way to really large global scale.

1:23Recently, you and Henrique shifted from co-CEOs for a long time to solo CEO. It's an uncommon arrangement. and I think a lot of founders would love to hear any advice you have on both that and then how the transition has worked. Yeah, so I think for many years we were the poster child of co-CEO success. So now it's changing that. Maybe we're just changing that pattern, I guess. So for us, honestly, not a lot has changed in how we actually run the company day to day. So we had this co-CEO structure where Enrique was more external, I was more internal. And the reality is, at some point, we've been doing this for really 10 years almost since our previous company in Brazil.

2:11And I think the learnings for us were, at some point, we massively simplified how we run the company over the past seven months, streamlining layers of management, priorities, fewer decision makers. And there were a few decisions that there was a need for more clarity on how decisions got made. and even though Enrique and I, I think, figured it out ourselves super well over the many years as the company scaled and we started to get a bit closer to a different level of maturity as a company, closer to being public and things like that, it became relatively clear that it would be good to go back to the traditional governance model of CEO and chairman.

2:53And we're happy about the change, but nothing really changed about how we run the company, how we handle customers, how we talk to investors or anything. So it's been mostly a formalization of what's been true for a few years now. It's interesting because when I look at European AI companies, there's an increasing number of them that actually have the co-CEO structure. And often it's a very technical founder plus a more business-centric or business background founder. And so that's companies like Helsing, Age. Like there's a couple of these actually. And it's some of the leading companies in European AI.

3:24And I don't know why it's so Europe-specific. Obviously in the U.S. there was Robinhood and others were, you know, that Ekoi structure as well. But I find it striking that in Europe, it seems like a more and more common pattern. So it's just kind of an interesting aside. Could you tell us a little bit about how Brex is starting to think about AI and some of the innovations and approaches that you're taking there? There's three big areas for us. One is the obvious, which is product. So how we can improve the product and make the experience of essentially expense management better. Right. And within product, there's two areas that we spend a lot of time.

3:57accounting is one and the other one is essentially expense assistant and expense management basically what eas would typically do second big bucket besides product is go to market and operations so things that are very ops intensive internally prospecting you know kyc underwriting compliance lots of use cases there and the third broad bucket is like developer productivity so how can help engineers be more successful. And there, I don't think we've done anything particularly remarkable in this third one because we're using the same tools that folks have used, co-pilot and things like that. We're experimenting with new tools, but I would say buckets one and two is where we spend the majority of time so far.

4:39Maybe even one step back from that, at what point did you say internally at Brex, like, I'm going to get up to speed personally on this or we're going to make this part of the product? Probably 18 months ago, I think right after Chachapiti launch, that we started to just play with, you know, ChatGPT online. And I had played with the GPT-3 through the APIs before ChatGPT was out. And obviously, it was impressive. But ChatGPT was that moment that everyone started to think, okay, what does that mean for my business now? The mental model that I don't think is particularly unique that we have is, you know, if we were to think about, you know, humans that are free, what would we do?

5:17It turns out there's a lot of work in expense management and accounting that people would automate, right? And the one that was really obvious to us early on is, you know, if we think about what is the best customer experience when it comes to expense management and corporate cards is essentially what executives have, which is there is no experience. You just swipe and it's done, right? There's an EA in the background that will figure out how to get a receipt for that, how to categorize its expense, how to get it approved, right? I prototyped something on the weekend in Python with GPT-3.5 on, can we just get someone's calendar and use that context to generate a memo, categorize an expense, and potentially find a receipt on someone's email?

5:57And the results were surprisingly good. And it took me probably, I don't know, 48 hours to get something working. And we tried it in a few transactions. And then probably in a week, we had a team try to build that on a side of Brex. And I think that was sort of the original impetus of us trying to get some of these things out because we knew that it could essentially give this pretty amazing customer experience for everyone. And today, we have roughly 30 ,000 customers on Brex that use Brex Assistant for expense compliance. And over a third of their expenses today are completed by Brex Assistant automatically, Which is basically something that would require them for going and manually doing.

6:41And, you know, over time, we want to get this number higher. But it's, you know, basically a net new product that we build by just playing out of a model in the beginning, seeing what works and getting it out there. That's really impressive velocity, especially in an area where, you know, I like from other fintechs to those sort of traditional financial services companies. A lot of them feel like they cannot use AI because it is probabilistic and there's like risk and reliability issues. Like, how do you how do you handle this or get something into production relatively quickly? So for us, that is sort of the holy grail of AI and thin tech, I would say, is like, how do you build this degree of conviction that what you're what you're suggesting is correct.

7:24Right. And and maybe maybe it's interesting to talk about accounting, which is one that people are fired if the results are wrong. But basically the way we thought of it is twofold. One is how do we expose ambiguity to the user, right? So instead of saying, hey, let me just try to predict something and put it in front of you and say, hey, this is what we generated. Good luck. We thought it would be a much better customer experience if instead of having, for example, like, you know, chatbots are particularly bad at this where, you know, a chatbot gives you an answer. There's no affordance. You can't understand other potential options that are generated.

8:07You can't understand context. So a lot of it is just like building ambiguity into the UIs and into the flows. And for example, like our expense assistant, when we're generating a memo, if we have really high conviction in a suggestion, we can go and say, hey, this is what we strongly believe is the answer. And we automatically apply it for you. If we're not that confident, we show you suggestions in the bottom of like a field. And if we're not confident at all, we don't show you anything. And then probably the second bucket, which is really interesting, is how to use traditional data science on top of these AI models to build good results, right?

8:42Because there was this phase when large language models came out where original data scientists were like, oh, my God, before I was the only one that knew how to do machine learning. And now everyone can do something that looks pretty good. And they were like, oh, what is my role in all of this? And obviously, I think as time matured, going from my prototype in my laptop to something that works across 30 ,000 customers and hundreds of thousands of transactions a month requires a level of statistical rigor and analytics that I don't think most people appreciate. So over time, those skill sets became really valuable in us scoring results from these models and understanding what is a good result and understanding how users rank a good result, which is different than what we believe is a good result also.

9:25So I think there was a lot of that of like solving for ambiguity into results and the quality of the results. And then the other thing that I think we spend a lot of cycles on is trying to think through how to add and contextualize suggestions that AI provides into a workflow that already exists. because I think, for example, if you're thinking about an accountant, right, I think very few people will be comfortable saying, hey, you know what? You used to categorize these expenses now by hand and close your books by hand. And now we're going to press this button and this bot's going to go in a corner and do everything.

9:59Very few people are comfortable with that. What people are probably more comfortable with is as we're going through a workflow, if you're categorizing expense and we say, hey, we noticed you did this same categorization three times. We believe there are 15 other transactions that apply to this. would you like to apply the same logic? Click here to review. And then it opens a nicely modal with all the expenses that are going to be affected and you can review. That's like a lot more palatable. So I think it's like the small things in like UI and these interaction patterns that go really far. So, and you know, I was looking at the metrics right before we jumped into this call, but we were looking at AI suggestions today and we have, using the strategy, we climbed to 80 % suggestion acceptance rate on the accounting side today.

10:46Wow. And I think if I started by saying, you know, let's go from, you know, click here, I'm going to close the books for you automatically, there wouldn't be enough trust in the platform to do that. And what happens over time is as people start just saying, okay, this is correct, I'm just going to click done. Then you can start, you know, building something that skips a step effectively. So I think it's something just that gets user comfortable with the level of ambiguity that these results can provide. Yeah, there's really interesting historical precedence to this too, where I feel like with each technology wave, there's things that users have to get used to and also new UI to help navigate them.

11:20And so, for example, in the 90s, people used to be uncomfortable putting credit cards into websites, right? The credit card number because they're worried it would get stolen. And, you know, there's other examples like that throughout history. and here you always hear people really worry about hallucinations or ambiguity and AI output, and it'll both get solved from a technology perspective. But to your point, there's really smart UI things that you can do. How did you think about where to allocate time from an AI application perspective, or how did you prioritize what to focus on? So for us, I would say the biggest driver was where can it create the biggest difference in customer experience?

11:55And again, unsurprisingly, we're pretty customer-obsessed at Brex. And for us, we were just thinking about what are the things that touch the largest number of people. And, you know, basically creating an EA for everyone that has a Brex card was a pretty high leverage initiative. And then obviously accounting is something that every single admin on Brex does every month. Everyone has to close the books. So I think it was very much thinking about just like impact and where can we lift the large number of hours in the system away from these folks doing these things manually. I think then the other sort of side of the brain where we went also was, you know, what does it mean for us internally?

12:42Right. Because at the end of the day, there's there's sort of two ways of, you know, cutting costs and automation from customers and from us is obviously, you know, directly putting that in the product and enabling customers to save time themselves. But also what are things that we can do internally to help us serve customers better? Right. And that's sort of the second big bucket of where we spend time on things like go to market, prospecting, demand generation and operations and compliance where a lot of the brain went. I would say my original impetus was very product oriented. And then right after most of the times where we spend time was on the marketing side, where we had a lot of sort of early thinking to do there, given that not a lot of people were actually thinking about how to apply this on a B2B setting.

13:31If you were going to look at sort of outside of the product to, as you said, go to market and the operations intensive parts of the business, do you have like a rank ordering of like where you think it's going to have the most impact? in general, scaling marketing is the biggest one. And the way I frame it internally to folks is like, you know, if you look at, again, the same framework applies everywhere, right? Which is just like, what would we do with infinite humans? And back in the day, if you think about marketing 10 years or 15 years ago, and you were at like Salesforce and you were trying to close, you know, Coca-Cola or a really large enterprise customer, you literally had an account-based marketing team, where you literally have a PMM, a product marketing manager, working alongside with a sales rep to market to that account, right?

14:15They're literally creating a pitch deck. They're creating materials. They're going to Coca-Cola and meeting the executives of very specific pitch of the value that Salesforce can provide and so on and so on. And really, the way I think about it is there is a world now where you can generate effectively account-based marketing for any accounts because the cost of doing that is marginal, right? So the way we think about this is like, how can we prospect across accounts that never got the level of personalization and that level of care with a lot of these tools and really, in a really specific way, provide value to these accounts in ways that we couldn't provide before.

14:58So, you know, for example, one of the things that we've done in the past is, you know, we're pretty, we started to this year to quantify literally how much value customers get from Rex. So, you know, effectively how many dollars would help them control spend, how much dollars they on budget, you know, how many hours they save, how many dollars they save, right? And over time, we were able to start being a lot more rigorous around, you know, folks in these industries, in these verticals, folks, companies that have these characteristics, companies that are global, for example, tend to have a lot more affinity and get a lot more value from Brex.

15:36So now we can say, hey, XYZ company, we know that your peer in your industry is actually using Brex. And they also have a lot of global employees just like you. And they don't pay any FX fees because they use Brex for doing this on a global scale. And this level of personalization is something that an SDR, for example, if they're just able to come up with. And a lot of the challenging part of what we're doing now is how do we build the systems that allow you to think of demand effectively as a system where you're basically tracking your entire TAM in your own systems and then effectively thinking how to target each account in a way that is highly relevant to the value they can get from Brex.

16:21And a lot of the content and a lot of the insights and a lot of the things that allow this to happen is effectively large language models because you can process not only a lot of unstructured data. So for example, is there any job postings about this company hiring somewhere else? Or even whenever I go write code email to an account, how do I write something that's relevant to them? And so on and so on. So that's a big area where we're spending a lot of cycles on now. And we think there's a lot of lift in terms of just creating a new playbook that hasn't been possible before. Yeah, I think that's really compelling because there are, and I'm not saying that these things are not valuable, but I think they are quite commoditized.

16:58There are many companies that are doing some version of like email generation that can plug into an existing email orchestration system. Yeah. And they'll call it AISDR and maybe they'll like plug into LinkedIn. Right. A lot of those. We've tried a lot of those. There's a lot of those. You know, what that means is there's very commodity and not very different from like one overworked SDR sending a spam campaign themselves, right? But what you're talking about requires like much more insight into like, what is a demand signal, a value signal, like specific data that Brexit itself has to collect because it understands that signal and wants to meet it now.

17:40And so I think like that should be much more powerful. A hundred percent. There's this book like Crossing the Chasm that is like a kind of classic go to market started book and talks about this idea of like, what is the definition of a market, right? and a very important concept is the idea of customers that can reference each other. So you're not really in the same market if this customer can't go talk to this other person about your product and hear something. Do they know someone that uses it? And effectively, I think now there is this ability of creating almost an infinite number of markets where effectively, I want to create a market of construction companies in Missouri that are high spenders on card or they have complicated accounting needs that may leverage breaks.

18:27And you can essentially outscale your ability of doing this with humans in a way that you probably wouldn't be able to do before. And to your point on a lot of the tools doing that, writing the email is the easiest part. That's actually not that hard. The part that is the hardest is aggregating all the data and essentially building this database of your whole team. Like who is every account in the market that could buy Brex potentially? And what do we know about them? And how do we enrich them with signal that doesn't necessarily exist out there yet that could give us a higher likelihood of converting those accounts?

19:09And I think the interesting thing there is, which I actually think is why this is really not positive, is ultimately your outreach becomes so much more relevant that people don't actually treat it as spam. Because at the end of the day, marketing is all about alpha, right? You want to be doing something that you are out, you're being more relevant than someone. You're being effectively more specific to someone's pain points. And I think the degree of personalization that you can achieve now when it comes to pain points, customers you know, understanding of your business and specific things that are going on in your life as a company, I think is materially higher.

19:48So I just think there is a lot of the work is the database and AI helps you actually build the signals and write all the email. But the actual hundreds of tools that help you compose emails of AI are not that relevant to this specific problem. One of the challenges, the thing you're talking about feels like a very, it's a pretty big technical effort, right, to do this data consolidation. And I think that there's a prior generation of go-to-market infrastructure that will become less relevant. Oh, 100%. Because it doesn't support like the scale intelligence of like, OK, I want all these first and second and inferred third party signals.

20:27The way I collect that signal has to be very conscious. I'm going to do it. I'm going to do intelligent analysis with LLMs in order to interpret that signal. I do think that it is a big strategic advantage for companies and then a big infrastructure project. And most many companies will be unable to do it internally. But it is an advantage for anybody who does build it. And we had to build it internally. I mean, we effectively build our own customer data platform for this because it's, you know, you just can't use Salesforce. I mean, the tools just don't really work that way. You can't ingest a signal in the way you would want.

21:03And really, like, a lot of the alpha comes from the edge cases, right? So, like, if you can just source this data through, like, Zoom Info, for example, and just put it on Salesforce, everyone will be doing it. The alpha comes from what are the signals that you can gather that no one else is looking at. And because no one else is looking at, no one is reaching out to those people that way. And therefore, it's extra relevant for them to do this in a really specific way. So I just think a lot of the marketing playbooks are being rewritten. And composing the emails or the ads is actually the easy part.

21:38It's just a signal that you need to even write a prompt that we spend a lot of time figuring out how to grasp. Yeah, I think one of the really cool things about Brex is that there's ways that AI can accelerate your business. But then there's also a lot of parts of your business that are very durable and robust in the face of AI. And I think these sort of AI durable businesses are increasingly valuable because you can't use AI to just directly attack what you're doing. Could you talk about the pieces of your business that are less impacted by AI or that you feel are really durable in the face of that?

22:06You know, people yawn at this. But, you know, when we started Brex, the number one feature people talked about back in 2018 is that you could go to our website, sign up and get a card instantly. And we were the first financial service company on the planet that allowed you to instantly onboard and get a corporate card. or literally a virtual card number that you can just swipe. And, you know, there's this sort of Jeff Bezos thing of, you know, what are the things that are not going to change in the next 10 years? And I think something that's not going to change is if we're still going to move money.

22:39And that is unlikely to go anywhere anytime soon. And, you know, there's crypto and, you know, we're bullish on that. And there's a few things we're doing in that space. But I think the fundamental thing is at the end of the day, especially when complexity adds up on the money movement side, that is the sort of underpinnings of all our software does. And the way I sort of explain our business to new hires at Brex is fundamentally the reason this business makes sense is because back in the day, you used to have your money moving on this side and that's where the banks were and all that. And then you have your software on the right side, which is where your controls work.

23:20And what we did at Brax is we just said, let's just do these two things in one. And when you do these two things in one, you have a lot more control and all that. And the software side, AI is going to change and make great. But the money movement is still challenging, especially when you get at scale. So one example is a lot of large public companies. And we just closed a$100 billion market cap public company a couple of weeks ago. Congrats. Thank you. Thank you. We have 130 public companies on Brex today. And one of the main reasons they signed up for us was because they need a card that works globally.

23:56They have, I think they operate across 20 markets, 20 countries, and they need to pay their card in local currency, settle in local currency. And the way we build our financial infrastructure, we don't rely on Stripe or on Marketo or any of these vendors. We literally go straight to the metal and do the money movement directly with MasterCard and local payment rails. allows us to do that at a really large scale. And that's the reason they signed up for us. And that has nothing to do with AI. And it effectively becomes a really good mode for us to then go and build a great software business on top.

24:28But these boring things are really a lot of the value. The other thing which is interesting, which relates to AI in this parts that are not necessarily as sexy as a software, is there are a lot of things there that are incredibly labor intensive. So for example, like compliance and understanding, monitoring a business and understanding whether that business has any compliance issues. So one of the things, for example, that we use a lot of AI for is adverse media monitoring. So if a company has any news that they're doing something that is potentially shady or illegal, we need to act on that. And companies have an obligation to monitor that, which really meant that companies didn't really do it because it was really hard to operationalize it at scale.

25:13And we actually have one of the most robust programs now, I think, in the country on doing that because we started to apply AI there. So even that part has some AI, but I would say at the end of the day, money movement is money movement. Maybe if we just like expand out from that a little bit, how do you think AI changes like finance even beyond Brexit's immediate plans? Yeah, I think, I mean, it's funny. We're still moving money, but you know is there is there impact yeah so so we we talk about this idea at brex of like there's like roughly three horizons in which we operate um and horizon one is like corporate cards that's where we started uh horizon two is what we call total spend which is this idea of how do we monitor every dollar spend even if it's not on card uh and bring that under our platform and and then horizon three is what we call continuous finance and really the idea there is like what would happen if you had this real-time visibility into every single dollar that flows in and out of your business.

26:13And effectively, the way I think about it is what finance teams are doing at the end of the day is a reporting job. You're reporting to shareholders, you're reporting to your business internally, and you're trying to aggregate and make sense of data that exists in all different sources in all different places. And a lot of the work of essentially accounting is a data cleanup job, right? You're basically getting data from these different sources and trying to make sense of it and categorize it. And of course, it's one that there are very high legal requirements and accounting requirements to be correct.