AI in Finance for CFOs to Stop Reporting Work and Drive Profit Decisions with Ron Nachum

Welcome to Future Finance. I am Glen Hopper, co-host along with my esteemed colleague, Mr. Paul Barnhurst. Paul, how you doing? Doing good. Always excited to be here. How you doing, Glen? I'm good, except I mentioned before we came on, I'm wearing a sports coat and a t-shirt today. I feel like Crockett from Miami Vice. It's a black t-shirt, so I, for those who are just listening, it's a black t-shirt, so maybe I like Crockett headed to a funeral or something. I don't know. We're live so we could poll the audience and see how many people even know who Crockett is or recognize him. That's true. I'm dating myself here by referencing an '80s television show that— Hate to break it to you, Glenn, but you and I are getting old. Yeah. I mean, we're both the other side of 50 now. Our next guest is not getting old. He's young and vibrant and everything that we remembered from all those decades ago. I wish I knew who Crockett was. See, I told you. Welcome on the show, Ron. We're really excited to have you here with us. Yeah, I appreciate you having me here. Excited to dig in. Yeah, we are as well. So I'll give a little bit of background about Ron. So Ron Natchman spent years in applied AI research before founding Sapien. Sapien is an AI-native system deploying agents to understand siloed, messy data to run, optimize, and act on company finances end-to-end. It sits on top of your systems of record, so ERPs, data warehouses, CRM, spreadsheets. Et cetera. Ron studied, if I remember right, see, was it Harvard or Wharton? Harvard. Harvard. Studied at Harvard. He's, uh, worked on several AI research projects and, you know, now he's, uh, scaling his company, which is powered by its company engine and tabular reasoning model. And we're really excited to get the opportunity to chat with you today. So again, welcome. Yeah, appreciate you having me on. I mean, great. Couldn't have said it better myself on the introduction. Perfect. All right, so my first question is really just an observation. I'm going to read you a list: Bill Gates, Mark Zuckerberg, Matt Damon, William Randolph Hearst, Robert Frost— you know where I'm going— and Buckminster Fuller. All famous, uh, Harvard, uh, folk who, who, who did not finish. And as I, um, think about that list, I'm wondering I mean, arguably that's probably rocket fuel for starting to becoming a billionaire industrialist, or at least being, hey man, I don't know. But tell, but there had to be, you know, parents were probably involved here, but it had to be a big decision point for you. What made you decide, I'm ready to go out into the world and do this thing and right now and leave school at that time? I will say there's probably a long list of folks who also left Harvard and didn't do so well. So I don't think we should just— no, no, no. That's it. Those are the only ones, you know, survivorship bias here a little bit. I don't know if we, you know, there's obviously a lot of others too. Um, jokes aside though, I, I spent my time at Harvard. I was studying computer science and statistics as well as a little bit of neuroscience. I was doing applied AI research. Um, I met a lot of really smart people and the way that I oftentimes try to do that is I was running my own. Things to just meet people. I started running my own AI paper reading group. I started a startup group. I was just like, how do you get smart people in a room in a way that's low touch? Just bring people together. It doesn't need to be a super regimented club. I actually considered dropping out of Harvard to join a startup at one point and decided not to for a variety of reasons, but it boiled down a little bit to the people. It was like, who was I going to be working with? That's actually what made this decision really easy for me. I started Sapien with two of my best friends from high school. Um, Pranav was over at Stanford, Aria was over at UT Austin. They are two of the smartest people I've literally ever worked with. Um, Pranav was like one of the best like infrastructure cybersecurity minds of our age group. He was like hacking things since he was in middle school and like got in a lot of trouble sometimes, but really did a lot of— any good stories on the hacking? I'm just kidding. I, not, not for it. Not for public disclosure. I figured you would say that. He would make me go and scrub it from the record here. Um, And so, so he was someone I knew extensively. And Arya also, like, he's awesome. He was a top 100 golfer in the world growing up and then went into computer science somehow instead of golf. Um, he ended up at a PhD research lab at UT Austin. He was their youngest, uh, student there. He was on track to get his PhD by like 22. So like two of the smartest people I've ever worked with. And to me, it all, it's about when you're building something, you want to be building up the right people. And so it was clear we had like the right talent density to at least get something off the ground and start building it. And then the other two pieces that kind of layered in afterwards is we met the folks over at, at NEO, um, Ali Partovi in particular, who runs NEO, the VC firm, and some of the other folks there as well. And it was one of those moments where you realize like the people aren't just the people in your company, it's the people you're going to be working with outside. And there was that moment when they first made us an offer to, you know, fund our pre-seed and kind of get the company off the ground. It was like, this is, these are people we want to be more like, like these are just people who are really excited about working with. So I was like, okay, I have the right people to build this thing with. We then really like started to nail down this space of like, look, financial data is such a big gap. We actually initially weren't certain exactly where we were going to take this, but it was so clear it's a humongous problem. And so to me it was like, look, I'm working with great people. We're going to work on a really hard problem that has high impact. And frankly, the world is moving so, so fast right now. Like the opportunity cost of not going and building something is so high. It made it a no-brainer decision to just go all in and do it. And, you know, now we're almost 2 years into the process from when we first started the company. And I mean, it's pretty awesome. Like, you know, team of more than 20 people now growing quickly. We're actually moving down the street to a new office in New York. Um, it was definitely a hard decision, at least for my parents was relatively easier. For my co-founders, a little bit of a harder time, but I think when you boil it all down, I think life is being around really talented people and working on really hard problems and When we got our first customer, we started to really see real impact. It's one of those things that made it so obvious, like, why would we ever want to go back to school when you can be, you can be, you know, saving thousands of jobs, you can be generating tens of millions of dollars for these companies, you can be having a really big impact. Um, but that's kind of how it all came together. Yeah, that's huge. And that's, I mean, and it sounds like, I think you're spot on with when you have those connections and those people that you have an opportunity to work with at a time like this is, you know, this is 1995 all over again. Well, probably 1995 on steroids. Uh, hopefully a better ending. Yeah, that's true. Well, I mean, you know, there's argu— I mean, there is obviously we're in some kind of bubble around it, but, um, you know, that also companies that started, you know, the Amazons that come out, so the best companies are going to shake out. And by starting building when you did, you know, you're getting— you have more more of a foothold to ride this wave. So yeah, totally makes sense. And I guess my second part, it's not even the second part of the question, it's a completely different question, but it's, I, listeners to the show know that I've spent 2 years railing against everyone calling their chatbot with some different kind of wrapper on it, and Paul knows where I'm going, calling it an agent. And this year, we are actually, you know, there have been huge leaps in how long the conditional loop can go out and work without a human in the, in the loop there. Tell me about when you first started building this, the idea of agents, how you were handling that, how much was left to the agent to kind of run on its own versus sort of an agentic workflow where you have deterministic steps in and I'd love to hear your insights on how much better the large language models have gotten and the, the AI tools at going out and performing these long run, truly agentic tasks. It's an amazing question. There's, there's a lot that's changed here. I'd say candidly, like our, our approach is very different than a lot of other companies in the space. Like there's been FP&A tooling, there's been analytics tooling, there's been all these things for, for decades now, right? Our whole angle was one where we were building an AI research team that really cared about financial and operational problems rather than, you know, an FP&A platform that was adding in AI. And what that meant is from day one, we were building this system we called the company— we called it the company representation. It's now kind of called the company engine. It's more, you know, user-facing, if you will. But it's this whole idea of like, why was this problem hard? Like, why did we have planning softwares that took 9 months to migrate into? Like, why were we even in this place in the first first place. It's because these systems didn't understand the business they were working with. And this will get to your question in a second, but it's really about, for every company, how do you build that underlying model of the business? And this isn't like a traditional Excel model. This is, what are the parts of the company? How do they relate to one another? What matters to this business? How does all the data, all these different pieces together? So the whole idea when we first started was just, it was just this piece. It was, if we can understand a business well, then we could build the most general company out there, right? We could work with a manufacturing company and a software company and a restaurant and a telecom company and healthcare business, and they could all leverage this one platform. So that's actually where we started. We started with this foundation of understanding of business. What that led us to initially is actually exactly what you mentioned. It was just automating repeatable processes. The first workflows we did for businesses, frankly, it wasn't even an agent yet. We were doing it manually in Excel in the background, just like trying to figure out like, okay, how do you put this together? Um, that was where we started and it turned into like, kind of like you mentioned, these structured tool calls that went end to end. But the, the interesting thing about it is we were always building a system where it was extremely open and you added in constraints. And so what that meant is as language models got better and as technology has been improving, we just kind of loosen more of the little dials and the constraints around it such that you let this company engine do more and more and more of the work. As you kind of models, models get better. What this has led to is when we started, the use case you were doing is like, you know, we were making waterfall charts for companies based on their GL data, but it was like very methodical. It was like, there's only one way to kind of get there. We then kind of went to Sapien generating entire management reports end to end for companies, still reporting, but much more in depth, a lot more thoughts there. And the immediate value is, well, you're going from a monthly surface level management report to. Can you get a mid-month, can you get a daily management report of the business that's going and breaking down to the customer level, to the SKU level, all those levels of detail? Then where we've seen Sapien go as these models have gotten better is again loosening those restrictions. Now Sapien is actually going into things like pricing recommendations, inventory allocation, like the actual, like, okay, we've gone from saving you time to a system that is saving our clients millions of dollars or generating them millions of dollars where they're losing opportunities. And that's the big picture vision for Sapien is because we built the technology in this way, we're not just building a chatbot, we're building this like system of action, the system of data understanding, the system of frankly what matters to a business. Sapien is able to then go and get proactive. And so where we've seen the most value today is we have companies that feed in every single piece of their data into Sapien. It's their ERP, their warehouse, their spreadsheets, their CRM, their HRIS, all of them are flowing into Sapien. And Zapier is sitting there and it knows what matters to the business and it's saying, well, I know you're trying to expand your profitability in, in these segments of the business. Here's actually what happened in the last 24 hours here and here's what you should do about it. And you're moving like less and less of needing the human to guide every piece of it. But what ends up doing is now every person is way more valuable. As we talked about at the beginning, now your people are not sitting there doing that manual spreadsheet stitching. They've moved up and up and up and up and up the stack to now it's strategic decision-making, executing on those decisions, going and making the actual changes in the business. And we've seen a massive shift both in how good models are, but also a lot of the work is outside of these language models, because frankly, LLMs shouldn't be doing most of this work, right? They're not built to do math, they're not built to do repeatable calculations. So it's all about how do you augment them with the right pieces. And I think we've kind of had the right timing where as LLMs are getting better, we can loosen up some of the pieces, but also our own proprietary technology, like our tabular reasoning model, which builds understanding of every column and row and table, our calculation engine, which does all the math. They all kind of ride this wave together where they're all getting better. So there's a lot of different pieces there, but I think it's been really interesting to see. We made this really big bet when we started that people do not just want a faster Excel or a better Excel. People want a system that helps them run their company better. And ideally you can also, you can export to Excel and keep some of the same ideas. That has definitely started to pay off as models have gotten better and better and better because this was not really even possible 18 months ago. But now we're seeing companies that are optimizing their business straight on a platform like this one. You already had the guardrails and the systems in place, so maybe, but I imagine, like you said, as the technology gets better, yeah, it, it enhances what you've already built. But have you seen anything in the way that you're set up where you're able to turn over more to the truly to let the LLM do the agentic work, uh, than you were just a few months ago? Or do you feel like, and we still need to keep these guardrails on, so we're gonna, we're gonna keep it locked down and on it, not let it just go rogue on us? Or— yeah, totally. I think, I mean, the way I would break it down is we definitely have less guardrails on the LLM than we used to in the past. Um, and I think it's one of those things where like you sometimes have to build stuff just to get rid of little bits and pieces as the technology is improving. But we try to think about, like, what are the parts that are— we call them, like, invariants. Like, what are the parts of the product that we're building that no matter how good an LLM gets, you're going to need them? One example is auditability. Every single number, you need to know exactly how it's calculated, all the way down to the source data, every assumption that went into it. And so we feel very comfortable spending time there. We feel very comfortable spending time on the data understanding layer and those sorts of pieces. One that we actually spend a lot of time on that has still proven to be really valuable is what we call financial intuitions. But it's just like the system having a gut feeling for like, hey, when you're comparing periods that aren't apples to apples, like, how do you think about that? Like when you're creating a table or a graph, like what is the way that people just like information presented? This is something that you would think would not actually be necessary. But when you ask, you know, any LLM to go and build you a board deck, it's going to get all these little things wrong. Like they don't make a huge difference. There's also the bigger things going to get wrong, but all these little pieces have come together. So I'd say we've let the LLMs be a lot more flexible, but we, we do see a lot of places where the work that we've done to build up these guardrails is still very, very valuable. And, you know, companies do their 10-Q reporting through Sapien. They do like processes that we would have never imagined being done through the platform. And that's only possible because you have these financial operational guardrails that make it structured enough. And I'd say that's actually the big weakness of LLMs in general is they're almost too open. Like, you can ask anything, they can give you anything. There's no governance, there's no control, there's no structure. Companies get a lot of unique value from Sapien because you're building a dashboard in one team and you can share it to the other team and they can go and drill in and get an understanding of it, but you still keep all the right, you know, structures in place. Um, but we've definitely seen it get more and more open, I'd say, over— especially over the last, you know, 4 or 5 months. There's definitely been a big shift with the last few models. Um, but there's definitely still a very healthy tension between the two that our research team spends a lot of time thinking about. I'm going to stop, Paul. I could keep going, but I know you— I want to ask more. You know, one thing when you said, you know, kind of the guardrails and they've got a lot better, but an LLM can kind of do everything, it made me think a little bit of what's been the, uh, what's made Excel so incredibly popular. But it also has been the bane to many people. It's incredibly flexible, right? You can do anything in Excel. You can do anything with an LLM. There's a reason I like to joke, and I think, you know, in some ways this is true with an LLM, a little bit different, but, you know, Excel is the number 2 software for pretty much every category in the world. Right, exactly. Right, you know, it's not number 1 for anything, but it's always going to be cheap. And if you, if you know enough about Excel, you can pretty much build it to be almost anything, at least at a small scale. You know, you could do pretty much anything with an LLM, but does that mean it's really the best tool to use? And so I think there's some similarities there in what you said. So anyway, that just kind of, kind of came to mind. There's one thing that I actually, I tell our team very often, which is I think in this day and age, the best products are opinionated enough to give you immediate value, but flexible enough to let you take it the rest of the way. And what that looks like at Sapien is. We onboard companies really, really fast because agents do most of the work to go understand the data. It's like, you know, multi-billion dollar companies get onboarded in 2 days, and it's like, it's really powerful what this actually looks like. The first thing they get when they open the platform is not an empty chat box. It's, hey, CPN's already gone and it's calibrated on your business. It's thought about your company, it's researched your competitors, it understands your space, but it also structurally understands your data. And you know what, we already found 3 SKUs that are gross margin negative, and we found these customers that you should be renegotiating with. That's the first piece of it. So it already understands that it's thought about it. That's the magic moment. But then everything is configurable. You can go and take it that extra mile and go and ask whatever question you want, but it's structured enough where it's going to still give you all the pieces of data that it took to get there, like all the pieces together. But I think this is actually exactly what I, I find the issue with Excel, with LLMs, et cetera, to be, is are they putting the power in the hands of the user tricky to tell. They're multipliers, right? It's how good is the person that using this, and then you can multiply them. I think the best pieces of AI technology are ones where it can— anybody can get value out of them, but the power users can get even more. And I think that's like a lot of what we think about and why these opinions matter a lot is can you build that piece of software? Like, we have, we have non-finance users use the platform a lot— operations, commercial, sales, etc. They're all using the same data, and it's a big boost for these folks to not have to go and ask their IT team every time they want to pull some data. But that means there's so much consideration for like, well, Sapien can't just show you a 300-line SQL query and say, do you think this looks good or not? They have, they have no idea, right? So there's so many things like that where, you know, it's that, it's that careful balance of like, do whatever you want with like, can you have enough of an opinion to not just be number 2? I actually really like the way that you framed it. It's something we think about a lot here, both on an engineering side, but also on the, you know, sales and how do we enable customer side of things. Great question to end on, Glenn. Well done. Continue to be Miami Vice. And until next time, thank you everybody for joining us. Appreciate you both. Thank you.