Fintech Corner Transcript: We Built a ChatGPT for Treasury and Finance

Joseph Drambarean (00:00:13):

All right. Big episode. We’ve been waiting for this one for a while. Welcome to Fintech Corner. My name is Joseph Drambarean. I’m the Chief Product Officer here at Trovata. I’m joined by Brett Turner, our Founder and CEO, fearless Leader, once again, hidden the trails to talk about how the world of fintech is not only continuing to be innovated in, but might be transformed in a pretty fundamental way by some of the technologies that we’re seeing and some of the things-

Brett Turner (00:00:42):

It’s been a cool couple of months.

Joseph Drambarean (00:00:43):

Yeah. It’s been fascinating. So why are we so excited? Chat GPT, the obligatory ChatGPT podcast, which has been a long time coming. We actually have been looking at ChatGPT for a while, so as soon as it came out, I mean, all of us were like, “Whoa, is this real? Is this something that not only could we use at Trovata, but is it going to be used generally in finance?”

(00:01:12):

We took a deep dive into looking at what does it know, first of all, because all of the examples on the internet were just so incredible, very niche examples, writing contracts, taking the LSAT, taking the GMAT and getting pretty much a passing score, being able to write entire essays on a topic. Your mind is just blown by that when you think of the creative possibilities, the outlets that you have with this kind of technology. 

So our question right away was, “Can it do math? Could you use it for finance things?”

Brett Turner (00:01:50):

Well, I think there’s the other aspect too of just it’s so profound in the way it just captured everybody’s imagination so quickly. Then I’ve got a lot of gray hair. I’ve been around for a while. So when something pops up, initially, “Okay. Is this going to be just a great way to maybe better curate cat videos or is this actually going to do really profound things like the internet was when it came about?”

Joseph Drambarean (00:02:13):

I think that the fresh wave of what happened with crypto last year, over the last few years and how hyped it was with all the coins and the NFTs, and I think we went through a wave of pessimism after that, but this is different. The difference is this has directly applicable use cases that today could actually transform our lives.

(00:02:37):

So when we started to look at this problem, and this has been a few months now in the making, so this is not fresh, hot off the press, we’ve been developing this hypothesis for a while. What we were frustrated with was the seemingly immense lack of context that it has. In terms of finance, what is the one thing, the one sin that you cannot commit in finance? Inaccuracy, right? You cannot make a mathematical error of any kind. There is no leeway for, “Well, that was close. You were only off by a percentage,” or whatever it might be.

Brett Turner (00:03:14):

Close gets you fired.

Joseph Drambarean (00:03:15):

Yeah. That percentage could be hundreds of millions of dollars. So that just doesn’t cut it. That was the first thing that we ended up seeing was breathtaking at the recommendations that it could offer. 

Especially, we looked at treasury, we looked at finance, we looked at FP&A as possible use cases where you could ask it open-ended questions about topics like, “How would I hedge a particular cash balance? What would the methodology be? How should I perform a forecasting exercise to understand my risk over the next quarter for a particular account?” complex question.

I mean, these are not easy questions, and it would answer them with incredible accuracy from a strategy perspective, but then when you give it the opportunity to give you the calculation, the net result, it would be so wildly off and you’d just scratch your head and be like, “Are you hallucinating?”

Brett Turner (00:04:15):

I remember early in the labs it was … I remember you showing me it literally was a table and there were numbers on the table. I mean, it was, I don’t know, a few columns and seven or eight rows. You could scan the numbers down and just eyeball it and say, “Yeah, that’s not quite right,” just having it sum up essentially the rows, which seems really right then and there, you’re just instantly, as a finance person, your credibility drop. 

Once you lose credibility or integrity of just doing the basic computational math, you’re done, you’re out, and that’s the way it is. That’s the harsh reality when you’re dealing in finance.

Joseph Drambarean (00:04:54):

We didn’t want to be discouraged because at the end of the day, it was providing value because it’s the equivalent of having another person, if you will, that knows how to do something. It’s just that that person failed every math course that they ever took in their entire life, but they just turned out to be brilliant in every other regard. 

So we went back to the drawing board to try to understand how could we use this technology and marry it with the capabilities we already have because-

Brett Turner (00:05:27):

Well, and one thing too, the big thing is context. What is it using to derive all that? This is all stuff that it’s being fed or it’s on the internet or things. So you think of this aspect of it doesn’t really know anything about Trovata. It doesn’t have any of the things that are … Essentially, this is what Trovata does. It’s automating cash workflows, but it’s not using any of those tools or have any of that data at its disposal, right?

Joseph Drambarean (00:05:56):

It maybe knows about us by our website. It knows that Trovata’s a cash management platform, but it doesn’t know anything about the customers of Trovata because we’re very serious about privacy and making sure that that data is never available. 

So then the question is, what value could it bring if it doesn’t have access to your transactions and balances? So we went back to the drawing board and we started to think about what if technology like this large language model technology, we could leverage the inherent benefits of its ability to navigate natural language because that’s the breakthrough. 

It’s the fact that you can converse with it and it’ll remember what you’re saying.

(00:06:38):

So just from a user experience standpoint, that’s breakthrough because it allows for operators and folks that might be looking at a tool like Trovata or any other tool, and maybe they have the cognitive load or dissonance of, “Man, I just want to do this one thing and I don’t want to have to poke around and figure out how to do it and maybe get it wrong,” and instead, they could just tell the thing, “This is what I want. Can you give me what I want?” The breakthrough is that it knows what you’re talking about. So how did we do that? I think that’s the question.

Brett Turner (00:07:16):

A big piece there, it was such a … I mean, a couple of things. One is, again, the whole aspect of the vision of Trovata is we’re helping automate. It’s not just initially people are like, “Oh, you’re the mint for corporates,” and it’s like, “Well, no, it’s not just pretty charts. It’s not just even better, more consumerized interfaces. It’s actually more of literally doing work for you, automating a lot of the workflows that you have, saving you lots of time, putting this power at your fingertips,” then the aspect of just even using natural language search and how to do that, whether it’s the categorization of all your bank data, being able to weave in other kinds of external bank data.

(00:08:00):

So the things that we’re already doing and leveraging, it just seemed to be such a natural fit because then it was like taking that that to a whole different level in terms of like a whole conveyor belt or a speed path. It’s just going to accelerate. So that even how we leverage search so much, I mean, natural language within already the interface, except you’re typing it in like you would search for something on the internet just for those little contexts with Trovata. So it just seemed like such a natural fit in terms of getting our hands on it too and playing with it.

Joseph Drambarean (00:08:34):

When we thought about this problem, and you make a great point because it’s more than just can ChatGPT answer questions, it’s Trovata provides functionality that extends beyond just question answering, right? It’s analytics, it’s forecasting, it’s capabilities that it might not be able to do without that context and without maybe even the capabilities at its disposal. So we just framed it a little differently. 

We asked, “What if instead of the end user having to be an expert in how to use Trovata, all of our tools, reporting, forecasting, tagging, all the capabilities that we’ve talked about, searching, what if ChatGPT could do that? What if it became the expert in our platform?”

(00:09:18):

Because it’s so great at understanding user questions, it would simply take those questions and filter it through its own knowledge of our platform and say, “I know how to find that. Let me go use Trovata to calculate the exact answer that you’re looking for and serve it back to you using Trovata and all of its native capabilities.” Then it becomes this hybrid where you don’t have to worry about whether or not the large language model is able to calculate something accurately because it’s not even going to try.

(00:09:51):

It instead is going to use a tool that is at its disposal and it’s going to use that to calculate, but then it’s going to use its own capabilities of context to be able to converse with you and say, “Well, you just asked me for your average free cashflow over the last 30 days and I presented it to you in a table, and now you’ve asked me, ‘Actually, could you put it into a graph so that I can see the trend?’ and it says, ‘No problem. I remember what I gave to you just a few moments ago. Let me put that into a graph.'”

(00:10:23):

That kind of interaction is really breakthrough because it feels very similar to the behavior you might have if you are on, for example, Slack or something like that and chatting with someone, a colleague, and trying to figure something out together, and you know that that colleague knows how to use software, knows where to look, has gone through schooling on finance topics so that you’re not having to explain basics. It’s as if you have an extra hand.

Brett Turner (00:10:52):

Yeah. Well, and I think that a couple things there, one, that translation layer is just so important because Trovata in the hands of an expert or a finance person, a treasury person, they just know what to ask. They know how to compile things. It helps them be a lot more autonomous in how they do that work. It’s that whole aspect of the robotics, of just generating or compiling or doing … So you’re doing that a lot faster with a lot of help, but you still have to ask the questions.

(00:11:22):

It comes from somebody who has the knowledge to be able to ask the right questions in the right way or be discerning enough or have judgment enough based on that background and knowledge to say when the result comes up, say, “Well, that’s not quite right. I know that right away because of this, this, and this.” Whereas somebody who doesn’t have that, they ask it, they get the results and they say, “Oh, great,” and then play that back right away without that filter.

(00:11:45):

So I think now when you basically take though that translation layer and start to automate that, then you really are bridging the gap. So somebody who is a business owner who may not have a lot of finance knowledge but they’re super smart about their business and they know the aspects about their business, there might be nuanced financial questions, they might not frame it up, they don’t understand gap accounting or revenue recognition rules or certain things, but they don’t need to because now you have that translation layer and that’s that last mile that’s so powerful now solved for you bridging that gap intelligently for you, and now it starts to make the dream come true a little bit, bridging that gap of closer to full automation.

(00:12:31):

I mean, it’s still early days, but the promise that we’re seeing when at the end of the day, not just using the academics in some ways and that powerful knowledge of ChatGPT and all of its computational aspects about it, but literally, it’s within the Trovata environment. So it’s using the tools, the ways to do math. It’s not relying on its own knowledge of how to do math that maybe it learns somewhere, who knows where, but now it’s using the right things that Trovata has at its disposal to be able to do those things. It’s just a marriage that’s beautiful and so powerful to see.

Joseph Drambarean (00:13:09):

That’s the other thing is that the underlying assumptions of where you got the data, how was the data cleansed, is this data reliable in the first place, those are handled by Trovata because let’s say you were to just put a bunch of data into ChatGPT, which you can do. You can take a CSV file right now with a bunch of transactions as an example, feed it, and it’ll process it for you and try to give you an answer. 

The problem is ChatGPT knows nothing about that data source other than you gave it to it and that because you gave it to it, it’s likely what you want. It doesn’t know whether or not it’s accurate. It doesn’t know which accounts it’s applicable to. It doesn’t know the context of how much data you have overall.

(00:13:59):

If there are, for example, missing transactions in that dataset, it would never know that. It’s not performing reconciliation for you to know if there are any issues, gap analysis, any of those things. So it is, in the world of finance, a garbage in, garbage out type of scenario, which is why we really had to solve this problem first. If you can’t trust the data, if you can’t trust the context, if you can’t trust the math, how could you ever trust this tool?

(00:14:31):

I think that’s where we have seen across all of the conversations that have happened in generative AI where, as you said, “Is this just another tool for generating really cool memes and pictures and stuff like that?” It’s obviously more than that, but the limitations are very clear as well where it starts to drop off dramatically, and it’s always context. It’s always, what can you expect it to know given the fact that you haven’t told it anything?

Brett Turner (00:15:01):

If you think of, “Okay. Is this going to fully automate a finance department or accounting department or a treasury department?” it’s more of if you think of what’s been going on forever or it seems forever now of just lean teams. 

If you look at the last 10 to 15 years, finance, treasury, accounting, everything is more and more compliance heavy and everything continues down this trend like we’ve seen in IT with cloud of just trying to get to cost compression, trying to get to more leverage and scale, and it hasn’t really made its way into finance and accounting or treasury just because there haven’t been the tools to be able to get there.

(00:15:42):

So when you look at these massive shifts with IT and AWS and how you’re saving 70% on your IT costs and things that have happened over the years, you just don’t get that same compression. What happens is you think you should. So there’s been a little bit. So then as you need to do more and the compliance rigor continues to add in the world that we have now, so when you think of this constantly having to do more with less, it’s incredibly difficult, challenging, and annoying for every finance, treasury, accountant because they’re having to figure that out. They’re left to their own devices to figure out whether their own ways to get a little bit of leverage, to do a little bit more, and then it becomes this vulnerability aspect of if there’s one little thing that interjects into this really tight schedule or right process that they’ve got dialed in, and it throws that off a little bit, you think of something like COVID and March of 2020, you think of everything dialed into this tight process and we have to do these things-

Joseph Drambarean (00:16:44):

There aren’t enough hours.

Brett Turner (00:16:45):

… monkey wrench comes in, you’re in trouble.

Joseph Drambarean (00:16:48):

You’re toast.

Brett Turner (00:16:48):

You’re toast, yeah. So everything has been compressed into the, “I’ve got to play defense so well to do all this thing from a protectionist standpoint.” There’s no room to really do anything offensive or getting this data out. So now, it’s like you have what could be the smartest person on your team or a team of analysts on your team. So if we could just get that one additional person on the team or if we could hire this one person and, “Well, there’s no budget for that,” now you’ve got something that extends your team and really solves that problem in a way that has been the recurring aspect or nightmare for every finance team out there.

Joseph Drambarean (00:17:34):

We can’t really overlook and stress enough how important it is to be able to meet that end user in the moment because if you’re stressed in that way, let’s say that rewind the tape all the way to COVID when it was happening and we were running around with our carts in Target wondering how many boxes of cereal we should buy and should we get the entire toilet paper section because we don’t know what to do. 

Well, the other folks that were reacting that same way exactly as you saying were in finance, trying to figure out, “How long can we keep our doors open? Can we make payroll if we have absolutely no sales volume over the course of the next few weeks? How many weeks do we have? What happens if this extends to longer than a few weeks? What options do we have from a liquidity perspective? We have lines of credit. What is it?”

(00:18:28):

Now, imagine being in that situation and every time you have one of those questions go through your head, the equivalent is hours of hard work in Excel to prepare data, to make sure that it’s accurate, to format it, present it, then make a decision, then go back and make another one and another one and another one. Imagine if there were tools at that moment in time that could keep up with the pace of your frenetic ideas where you’re trying to get answers quickly and you just want to tear down the barriers.

(00:19:06):

You might not need perfection in terms of formatting or the exact colors or whatever you might need on the visualizations, but you need answers and you need to trust them and you need to build on those Lego blocks of answers because if you are able to get to a conclusion, maybe that means that you do make payroll and you found a creative solution, right?

Brett Turner (00:19:27):

Well, and you’re talking scenario planning, and that’s just one of the hardest things to do. That’s why people who are really good at building, and there’s a whole art to constructing models in Excel and being really, really good at it. 

Everybody can build a model, but there’s few people that know how to build them really, really well. If you’re really trying to build tons of leverage in there so you can put certain inputs in and then it’s going to automatically calculate or drive a lot of, from a set of assumptions, what those results are going to be. 

Those are going to be things that are going to really drive value or impact or answer big questions or they might be at the point of making a pivotal decision for your business.

(00:20:06):

Those are game changing things, but if you don’t have it built well or even really good models, you’re putting in assumption and then it’s like you’ve got to turn the crank so to speak. It’s this big machine and you’ve got to figure out, “We’ve got to spend hours and hours with lots of people to figure out, ‘Okay. What is that going to translate to in terms of action that we can take?'” That goes on. That’s been going on for so long, decades, forever.

(00:20:30):

So if there’s truly a way to be able to feed assumptions in, but that whole turning of the crank isn’t done through these clunky models or risk with potential error, and it’s done at 2:00 in the morning, this is going to determine, like with COVID, a lot of finance people or treasury folks weren’t able to go and get groceries or the toilet paper or the cereal. They were locked in a room for 48 hours trying to figure out, “Is our revenue going to complete? If our revenue falls off 30%, are we going to be able to stay in business or are we going to have to make a really tough decision now to let go 20% of our workforce or 50% of our workforce?”

(00:21:13):

I mean, these are affecting lives, huge, monumental decisions, and we all went through that phase, which was terrible. So you’re thinking all these ominous things. So you feel this pressure and then you don’t have great tools to be able to get those answers quickly, and then you’re just pressed saying, “I hope I didn’t get …” or you’re rushing to build something and, “I hope there’s not a formula error,” or, “What did I miss?” and you’re not getting enough eyes on. I mean, that’s all the same stuff that’s been happening forever. So if you can have something that really helps with that process, I mean, that’s just holy grail stuff, but it does require a marriage of a lot of great things together. What do we got?

Joseph Drambarean (00:21:56):

Why do we have a laptop here? For those of you that have watched the show a few times now in our studio, you’ll notice that the laptop has replaced our plant.

Brett Turner (00:22:05):

Where’s our plant, our friendly plant?

Joseph Drambarean (00:22:07):

Maybe really quick to set this up too, let’s talk about what really we did in the last couple of months. I mean, that was really the task. It can’t be something gimmicky because the other thing is everything we do is great intention with great … I mean, Trovata, trove of data, we’re building the vision, the most powerful trove of financial data on the planet to automate workflow. So everything we do, it has to be done at absolute integrity. That’s why we’ve actually started working from bank data and built APIs with the banks because that truth of record is so important in terms of the underpinnings of cash flow everything is built off of. So we don’t let you put your own data in. If you have your own bank statement and put it in, we don’t just let you do that. It has to come through the API or directly from the bank to keep that-

(00:22:59):

We’ve been trusted, right?

Brett Turner (00:23:00):

Yeah. So if you think of now what exactly have we done here in the task of making this truly valuable for Trovata users or for companies, it’s the marriage of some really, really cool things that have been lacking in finance or fintech. Maybe this was the big task of, in our labs, let’s make this all happen.

Joseph Drambarean (00:23:22):

I’m going to get a bunch of the credit to Francisco, who had the vision to see the possible with this technology because I think that our reactions, a lot of the folks that saw this for the first time was exactly that it’s cute, it’s interesting, but it’s not ready for primetime. It needed someone to see the possible and forecast into the future. If we could solve these problems, then it would be a breakthrough. It would be something that everyone would want to use, especially in the world of finance. I think it’s-

Brett Turner (00:23:56):

Shout out to Francisco, our VP of Machine Learning.

Joseph Drambarean (00:24:00):

Amazing work.

Brett Turner (00:24:01):

I mean, this challenged him in a lot of ways, and then he embraced it. All the stuff that he was working on, he’s like, “Okay. Set that aside. Brace this,” and he went to work.

Joseph Drambarean (00:24:11):

Absolutely, because it’s the type of thing where the first time that you held an iPhone in your hand, if you can just recall those tingly feelings of like, “This is different. This is going to change things,” and we are starting to feel very similar feelings where it’s a frontier that is evolving so fast and is so relevant to our space and to every aspect of technology that you almost feel this gravitation, this pull towards that you have to figure out a way. I think that that’s why his willingness to work through these problems, and I think of three pillars here that were really important to the approach.

(00:24:51):

Number one was the quality of the context. We’ve been talking about that a bunch, but can’t stress that enough. Without that, this is nothing, right? You don’t have anything, accuracy, the focus, the maniacal focus on, will this be true and can you expect it to be true? Then the third pillar, which is-

Brett Turner (00:25:11):

Got to be able to bank on it.

Joseph Drambarean (00:25:12):

Right. Exactly. The third one though that we haven’t talked about at all is privacy. Can I trust this thing? Will I have to share every little piece of information about myself and our finances in order to get the value that you’re talking about? Those three had to be in concert together to have a viable solution because this is not working in social media or in another arena where you might say, “Yeah, go ahead. Get access to all of my tweets. I don’t care. If it means that I’m able to generate better ones, that’s awesome,” or if I want to use my social posts in some other platform. It’s very different when you’re talking about publicly traded companies, you’re talking about highly sensitive information that cannot be shared under any circumstance with anyone, even for the sake of automation.

(00:26:08):

So we had to find a way to take the inherent benefits of the large language model approach of understanding language and being able to react to it and have context while also getting the bedrock security privacy that you get in Trovata, which when you think about it, how did you do that? Well, the way that we had to do that is by what we talked about a moment ago. We had to teach ChatGPT how to use Trovata in a safe way, in a way where the user is in control, where we are not sharing any data with ChatGPT in terms of your transactions, your balances, but still at the same time giving you the ability to have that context so that it can in fact process that data and work to get an outcome.

Brett Turner (00:26:55):

Well, essentially, would you say it’s just a private version of ChatGPT? We’ve allowed ChatGPT to come into a very private, closed loop setting, and then it’s basically using all of the things at Trovata at its disposal from data and from tooling.

Joseph Drambarean (00:27:12):

So we’re not hosting ChatGPT. It’s not running on our servers. It’s still what it is. Open AI owns all of that technology. It’s theirs. What they don’t have access to though is our API platform, our user’s data, the context of how to use our platform, and what to do within it so that you could actually get benefit out of it. I think that’s where the magic is happening.

Brett Turner (00:27:34):

Let us calculate the math. Don’t let ChatGPT calculate the-

Joseph Drambarean (00:27:38):

Let’s work together, right? If there’s a fence, you can’t cross that fence in terms of privacy, but we can send messages over it. If I say to you, “Hey, user just asked for what is an FBAR report of my last year’s worth of data across all my accounts, can you produce that from me in a table?” you can send that information to ChatGPT, and that doesn’t say anything about the user’s data, right?

Brett Turner (00:28:06):

Well, and that was one of the cool things that happened. We don’t actually have that built into our platform. Honestly, we should, but we don’t yet, and it’s more of a canned report, but it’s basically just IRS guidelines that say, “Here’s how you …” It’s like filing your 1099s or it’s, “You need to file this report. Here’s when you need to do it, April 15th.” So it provides the IRS guidance. It basically knows that guidance because it has that indexed, and so it was able to basically just build that report on the fly, which was incredible.

Joseph Drambarean (00:28:38):

That’s what’s so cool is that FBAR, unless you go and Google search it, which most people would have to to even know what that acronym means. It is a hyper niche piece of information that is really only relevant to a very small part of the finance workforce, the treasurer.

Brett Turner (00:28:56):

So it’s basically financial accounts, the reporting of that that you have to do because the IRS and the US government is looking at international companies and some companies that’s got it on the watch list, it’s making to, “How are you doing business internationally?” So it’s usually just for larger companies, generally. So in treasury, everybody knows about it because treasury, the function exists usually for larger companies that are doing business internationally, not all, but generally. Some finance teams that are big enough and doing business internationally maybe that don’t have a treasury team yet will do that, but it is a little bit more of a bigger scale, but you could be a smaller company in doing business internationally having foreign accounts, you’re subject to that.

Joseph Drambarean (00:29:41):

I guess the other concept here being because it’s niche in terms of the folks that would ever have to produce a report like that, the knowledge of it is where we see the breakthrough because ChatGPT knows what an FBAR report is. Why? Because it’s read every possible website and PDF and book that has ever been written about treasury. So it knows that-

Brett Turner (00:30:05):

The irs.gov.

Joseph Drambarean (00:30:11):

Irs.gov, right? So it knows that an FBAR is this, it looks like this in terms of structure, it’ll have these columns as a requirement, and the data that you need in order to populate it is this. So we toss that over to ChatGPT. It looks it up. It says, “Oh, I know what an FBAR report is, and you are saying that you need it for the past year. Great.” Well, now, how does it come back and get the data?

(00:30:35):

So this is where the magic is happening. It’s because we have formulated an approach that gives ChatGPT just enough context about how our database is structured so that it can then go and query it using our APIs, granted the privilege to do so using the user, but it’s not receiving the data. It simply is formulating a perfect request for how you might find that data in our database, giving it to us, and then we execute it. Because this experience is in our platform, that data never leaves. It gets presented inside of Trovata. It gets visualized inside of Trovata, and all that we ever really received from ChatGPT is the instruction set. It’s, “I know what an FBAR is. It’s made up of this type of data. Here’s the query that you would need to follow in order to arrive at this conclusion.”

Brett Turner (00:31:34):

So would you say this equivalent is almost like … So ChatGPT being this genius or a team of really smart financial analysts, they come to the military compound, if you will, and we’ve essentially given a badge or security clearance, we’re making sure that they clear through the gate, they’re not bringing anything with them or any of that, they’re fully frisked. They go through that. They come into the secured arena here and they go into a room, they sit down and they’ve been given tools, they’ve been given access to the data. They can do all these things, but when they leave, they go out in the same way that they came in.

Joseph Drambarean (00:32:17):

Is that analogy? I would only modify it slightly by saying imagine that it walked into that room and it’s been given access to people that are sitting with their hands on the buttons for the tools, but it is not allowed to touch those buttons. It can simply say to those people, “Hey, I need you to run this radar over here and then whatever you do, give that to that person.”

Brett Turner (00:32:42):

Control aspect. Yeah, that’s a good point.

Joseph Drambarean (00:32:43):

“Just do this. You do that, you do that, and whatever I say, it will be the answer.” It gives it this … We’re taking advantage of that genius, its ability to know the answer, but we’re not actually letting it manipulate the data, touch the data, control the data, do anything with the data, really. Why are we doing this? Why are we so serious about this?

Brett Turner (00:33:07):

Trovata being that buffer, that control layer of protection essentially to keep that guarded.

Joseph Drambarean (00:33:13):

The reason we’re doing it is because our end users, our customers trust us. In the first place, they signed agreements with us. We have said to them explicitly, “Your data will never leave Trovata. It will never be accessible to anyone outside of your organization. We’ll make sure that we follow and pass every possible compliance rigor you could imagine for this type of data.” We can’t seed any of that in the interaction here.

(00:33:42):

So what we think is such an important aspect to our approach is that you get to take advantage of that genius in ChatGPT, but you get to keep everything else in terms of that privacy, security, all of the peace of mind that Trovata is controlling all of those pieces, and the result is incredible. We have it here on the laptop and I’m sure that we have it on the screen right now for users to see, but the net result is ultimately an experience where in Trovata, you have an interface where you can interact with ChatGPT. You’re not going to ChatGPT, the website, the Open AI, you’re not going to Bing, you’re not going to any of these websites to interact with it. It happens in Trovata in a private session that is really only available to you as a customer.

(00:34:34):

This interaction is not shared with other customers. It’s not shared with any other user. Your data never leaves Trovata. It’s not available to ChatGPT as log history, chat history or anything like that. It’s private. As a result, as you can see here, you can start to ask really complex questions.

Brett Turner (00:34:53):

Yeah. So maybe walk through a couple of the basic things that we’ve done. We’ve done more complex things and it’s so cool to see just the iterative nature of it and how it’s going to continue to build on itself, but maybe just a couple examples of-

Joseph Drambarean (00:35:06):

On the screen right now, we have a couple of interesting prompts where one of the interesting things that you might want to know if you’re doing an analysis of your transactions over a period of time is which specifically out of those transactions might have been recurring and might have been large in terms of the overall amount in your cash flow. Why would you need to know something like that? Well, if you’re trying to provide an analysis of where you could cut spend, you’d want to start with where is most of the money going out. The second element of that is if it’s repetitive, if it’s happening in a way where it’s predictable, then it’s a key-

Brett Turner (00:35:43):

Like a subscription.

Joseph Drambarean (00:35:44):

Right. It’s a key opportunity.

Brett Turner (00:35:47):

Which is a growing issue. I mean, people have so many subscriptions to so many different tools to be able to say, “Hey.”

Joseph Drambarean (00:35:53):

That’s exactly right.

Brett Turner (00:35:56):

“Give me top 10 tools,” and then you can quickly ask, “Are we even using this tool anymore?” and you can quickly see what that expense might be.

Joseph Drambarean (00:36:04):

Exactly. Imagine a CFO just being fed up, “Every time I hear that there’s a team buying another tool, they haven’t gone through procurement. Does anybody follow any process? I want a report.”

Brett Turner (00:36:16):

Actually, it’s happened with me and it was probably a few months ago. For us, we use so many different tools. I went on a little bit of rant, but it took a bit of work to compile all that. I mean, this is something that’s compiled in seconds.

Joseph Drambarean (00:36:29):

Now, imagine, this question, it had a couple of interesting keywords in it. It had largest, recurring. It’s looking at a dataset transactions. Then you’re also saying, “All of my accounts, not just some, all,” and you’re giving it the specificity of, “I want to know this within a range.” So imagine all those pieces of context are being derived just from you typing a sentence, and then it takes that and formulates it into a table. I didn’t have to ask it to do anything. It just said, “Oh, no problem. I couldn’t find those for you,” and then it gives you that output.

Brett Turner (00:37:04):

This is one of the reasons why, again, just the vision behind even starting Trovata and a little bit of my journey as a CFO of looking at getting closer to bank data and just true bottoms up cash flow, cash forecasting, cash flow analysis is because it is really hard to do analysis from data in your ERP system because you’re dealing with something that’s all predicated on accrual-based accounting. It’s all following gap and accrual-based accounting as well. So you have a lot of these accruals or estimates that you’re making month end, and if you’re closing the books really fast like a two or three day close at month end, then you’re making tons of estimates of what those expenses are.

(00:37:43):

So when you try to do even your analysis from ERP data, you got to then sift through what’s the estimates, what’s real, “I just want to see the cash outflows,” it’s really hard to pull that out. So when you’re able to do it just strictly from the bank because there’s no estimates or accrual-based accounting, it’s straight, this is what was paid, this is what is expended. So as an FP&A tool, I mean, it just allows it to take these kinds of analysis FP&A wise to the next level. I mean, we’re already doing that, but to now get this on the fly recall is incredible.

Joseph Drambarean (00:38:18):

Imagine, we didn’t have to tell it, “Hey, when you look this up, I want to have a little bit of context so that when I do my analysis, I have what I need in order to make a decision.” It just figured that out. It gave us a column that gave us the transactions count so that we could say, “Oh, interesting. There have been 17 occurrences of this really big transaction that is happening from a specific account and it’s in USD currency and here’s the average amount.” It computed all of those things without me asking anything of it, right?

(00:38:53):

This is the thing where because it knows the standard operating procedure for a lot of these questions, it’s taking from the inventory of all of the reports, all of the finance assumptions it’s ever read about and it’s saying, “Given what you asked, I think this is what you’re looking for.” This is why we get excited is because let’s say that you were to text someone on your team on Slack and say, “Hey, do this analysis for me,” with that prompt, one sentence, not very much context, it’s just, “I need 30 days. I need it to be recurring and I need it to be across all accounts and it has to be the biggest ones.” Would they know that specifically, “I need you to also show me how many times it occurred. Do the extra work. Go the extra mile. Because I’m not going to just trust you that you’re going to find the biggest ones. I want you to just show me your homework.”

Brett Turner (00:39:47):

Well, in the real world too, when you start … Let’s say a CEO or CFO is asking somebody on his or her team various things, there’s this interaction that needs to happen. It’s part of the relationship. You need to know the context of what they really want because you might come to the table with what you think they want and then you might be a little off and they say, “Oh, no, no, I need this and this.” Pretty soon though after a few months or it could take a little longer, a couple of days, you’re recalibrated. You’ve learned what that person wants to be able to feed it up in the way that they want to see it, but that takes some time.

Joseph Drambarean (00:40:22):

Absolutely.

Brett Turner (00:40:22):

So this is where the power of machine learning, also the things that we’re doing, helping this along its way to accelerate that journey to know really quickly who is asking the questions and also how to put that data together in the right way as well.

Joseph Drambarean (00:40:37):

Another little piece here that is worth thinking about is in order to generate the transaction count, the average amounts, getting all these insights, well, you have to have the analytics capabilities to do that in the first place. If you didn’t, how would it be able to arrive at that conclusion? These would be empty columns ultimately. So there’s the marriage of it’s super great at understanding context, but if it doesn’t have the context and it can’t calculate the context, then it’s worthless.

Brett Turner (00:41:06):

I mean, it’s a perfect marriage because this is what we’re built. I mean, as a fintech, we have the largest library of APIs with banks. We have this massive and growing trove of financial data, of bank data, and users are curating their own financial position across their accounts, whether a foreign currency exposures, they’re categorizing all their data, they’re letting us through auto tagging, categorize some of their data.

(00:41:34):

So the data management aspects of that, I mean, that’s at a core how we’ve built this thing. That’s the fun of building block of Trovata because everything you can’t automate, if you can’t have the data underneath it all locked and allowing that to happen. So it’s leveraging all the stuff that we’ve built.

Joseph Drambarean (00:41:55):

So let’s keep going on your point, which was the CFO had … Maybe it was a junior accountant or someone looking into some of this information. In your example, they have to go through a little bit of training, and I’m sure that every time that it happens is there’s a grumble as they’re walking out of the office saying, “Of course, you want it in a different format. You couldn’t just take my original.” Well, what happens then? They have to go back to work. They have to go through that data again and you might have to wait a few hours to get the output and the customization.

Brett Turner (00:42:29):

I really have to put it in a pie chart or-

Joseph Drambarean (00:42:31):

So what happens if you say to this capability, “This is good, but, hey, can you plot this in a bar graph for me?” Well, that’s the cool thing is that it can take that kind of information and two things are happening. Number one, recall. It remembers everything that it said to you. “You asked this question, I produced this data. I got the data in this way, and I also collated it so that it had this formatting.” Notice the prompt. We didn’t give it any more context. We just said, “Yeah. Can you just give this to me in a bar graph?” So imagine you’re the CFO, you’re like, “This is good. Can you put this in a bar graph for me?”

Brett Turner (00:43:13):

It’s like, “Done.”

Joseph Drambarean (00:43:14):

Yeah, and it’s like, “No problem. Here’s your data and here’s the other important piece.”

Brett Turner (00:43:19):

Well, the CFO, they’re saying, “Can you do it in the next few minutes? Go back to your desk,” or, “Send it to me in a couple hours.”

Joseph Drambarean (00:43:27):

Yeah, and, “I’m watching.”

Brett Turner (00:43:28):

Yeah, “Can you do it right now?” I mean, “Done.”

Joseph Drambarean (00:43:32):

What’s cool is that, again, there are a lot of micro-decisions that are being made about the formatting of this information. The X axis in this case is taking every single one of those recurring transactions, itemizing them, and grouping them by the largest ones so that you can do an immediate analysis of, “Oh, the ones on the left are the main ones that I have opportunity in.” Then check this out. It also is smart enough to group it by currency because it looked at the list and it said, “Oh, that’s interesting. There is a natural grouping that is occurring here, so I’m going to make a different color choice for the Canadian dollar transaction versus the USD ones,” and it did this on the fly. You didn’t have to tell it anything.

(00:44:21):

The context was, “Just put it into a bar graph,” and I think this is where the breakthrough is is there’s this trustworthy counterparty, someone that you can just … It’s like you can dance with this partner and trust that if I give you this open-ended, almost nonsensical question, if I gave it to someone else, you can take everything else that I’ve ever said, know it, and be methodical with it, do something with it. This is why we’re, I think, getting really excited about this. It’s because at the end of the day, it’s a completely different interaction pattern. We have all kinds of interesting ones.

(00:44:58):

So a common use case is identifying anomalies. The previous example, we were looking for an easier exercise. One, what’s the recurring transactions? In machine learning, it’s actually a complicated problem to do because what’s considered a recurring one, what is the calculation that you’re using. Is it time? Is it recurrence of certain types of descriptions? Is it both? Is it a certain period that you should be looking at, 90 days versus 30 versus whatever it might be? So it’s taking care of all of those complexities, but the next step is, what about anomalies? What about things that are outside of the bound of-

Brett Turner (00:45:35):

Which is all the stuff that scares the heck out of everybody.

Joseph Drambarean (00:45:38):

Right. Exactly.

Brett Turner (00:45:38):

Especially when you’re dealing with lots of banks, lots of accounts, lots of data, it’s just impossible that your team is ever going to pick that up, and you know something’s in there that you haven’t found or you don’t know about and you’re afraid when it gets unearthed because it probably could be pointing to something that could be fraudulent, potentially, and you just don’t know.

Joseph Drambarean (00:45:59):

So let’s reimagine our use case here. Maybe you’re the CFO and this time you’re driving and a thought pops into your mind of, “Something about our balances is off, and I have this sneaking suspicion that there is a transaction or a set of transactions that snuck through and we didn’t catch it.” So you pick up the phone and you call maybe your controller and you say, “I’m going to need you to look into this. I want you to do a deep dive, go through our bank statement, find every single transaction that happened in an out of bounds way. Find them, give them to me because I need to review every single one.” What a nightmare task. You have to then download every single bank statement. You have to format them, get them ready, and then you have to decide, “Well, what is considered an anomaly?”

Brett Turner (00:46:52):

Just pause right there. Most companies don’t even have their bank data catalog. Again, this is why working with a little bit more mid-market, small enterprise companies, but you move up market, large companies. If you’re using these legacy treasury tools, it can take years to connect all your accounts. So just getting all your data in a normalized fashion so you could leverage it, let alone even historical data because even the legacy systems aren’t great at even having all that catalog where you can utilize it on the fly in ways like this, there’s no chance. Then most companies that just don’t even have it, they’ve got to then start downloading historical CSV files from their online banking, there’s no chance, right?

Joseph Drambarean (00:47:36):

This question is ultimately not a, “Hey, I’ll get back to you in a few hours.” It’s, “We might have to do a deep dive. We might have to take a few days to do this, and we’re going to have to double check our work because at the end of the day, what’s the methodology that we’re using to identify that anomaly?” It’s not just as simple as, “Oh, this one was a standard deviation of one on the amount of every other transaction.” There could be other things. What if it was an anomalous count? It happened more than five times over a period of time and it was unexpected.

Brett Turner (00:48:07):

This also speaks to the why we’re working closely with banks. We have banks that are investing, our investors in Toronto like JP Morgan, Wells Fargo, Capital One, but you look at that. If you think of they’re not tech companies, it’s hard for them to build something nimble off of the data that they have, they have the data. We’re doing that through the APIs of getting the data, but if you look at something like anomaly detection, that should be standard in every single bank portal. You should have something that if there’s something anomalous that sounds out from the masses of how your bank transactions, the bank should have that just builtin, just scanning and just provide and tee that up for you in the bank portal.

(00:48:49):

So these are things where the bank portals aren’t really able or the banks aren’t able to keep up with just basic functionality. It’s like when you step into a brand new car, things that are just there now. You’re not doing this for the window and all these basic things that you take for granted like sensors on somebody’s next to you coming up on the next line, but this is why you have to have these modern things. Everybody expects when you’re backing out of your parking spot that you’ll hear the beep, beep, beep, beep. It’s a sensor. I mean, that now is it should be standard in every single car. Everybody now takes for granted.

(00:49:28):

So I don’t know if you’re driving in a newer car, you’re so used to that and expecting it. If you drive a friend’s car or your parents’ or an older car and it’s not there, you could literally hit-

Joseph Drambarean (00:49:38):

You bump into a pole.

Brett Turner (00:49:40):

Yeah, exactly, but this is where all these things bringing … They really should be standard. Everybody should have these tools and they don’t, which is crazy.

Joseph Drambarean (00:49:48):

Right, and what happens here when you look at it is that once you get used to this kind of behavior, everything changes because in our example, let’s say that you made that phone call, they did the research. Well, the next question you’re going to ask is, “Show me your homework. You found one and I’m scared now. How did you figure that out?” We asked it a very simple question. It found an example of a transaction. It’s actually a pretty big one. So now we’re scared and we’re asking why, “Why was it anomalous? Show us what you did,” and it replied with, “To determine why a transaction is anomalous, we can compare it to the average transaction amount for the same account and description over the last 30 days, and here’s the comparison.” It gives us that exact comparison, what was the average amount and what was the max amount, and that’s why it identified and flagged this one, and it was for Canadian dollar transaction.

Brett Turner (00:50:41):

That’s amazing because at the end of the day, when any finance person is putting together something, an analysis or report, even something that ends up might being on a slide that gets presented by the CEO in a board meeting, at the end of the day, everybody wants to know, “What’s the lineage? What are the assumptions? How did you do your work? I just want to know a basis for what I’m looking at in that context,” because there could be a couple of pivots or things there. I mean, to be able to get almost like a log, “Here’s what I did, here’s how I did it, and here’s what you’re seeing as that context,” because you could quickly read in there and like, “Oh, wait, no, I wanted 90 days.”

Joseph Drambarean (00:51:19):

Which you could turn around and ask, right?

Brett Turner (00:51:19):

You could turn around and … Yeah.

Joseph Drambarean (00:51:24):

Obviously, we could go all day. We could spend hours on a podcast to show the different examples and why they’re relevant.

Brett Turner (00:51:31):

This won’t be the first podcast.

Joseph Drambarean (00:51:33):

For sure, And we’re choosing examples that are very contrived right now, obviously, because they’re things that we’re excited about and we know they would be difficult to do if you just wanted to build it as a feature. Thinking of what we went through recently and other businesses went through recently when there was a scare with SVB and everything that went on and thinking about cash flow and understanding cash burn and going through that, there are so many platforms that claim that they do cash burn analysis and how you might do that, and we actually do it as well, but one of the things that we were curious about was, “Could a tool like this help in a moment of fear like that? What does my cash burn look like right now? If it keeps going, what will happen?” Being able to ask those types of questions and getting immediate analytics, right?

Brett Turner (00:52:25):

Trends are, I mean, just the hallmark of always getting early indicators, just even basic things that even pop up in a board meeting. Over the last few startups, there would be inevitably some … If you have a business that has inventory or an inventory component, one of the signs in any inventory, a business that needs inventory is inventory turns. So that’s always an indication. Every analyst looks at that because if the inventory starts to go down a little bit, you’re not turning that over quick enough. It’s starting to get a little stale. That means your sales are slowing down.

(00:52:59):

Maybe not necessarily. There could be some other, but these are all early signs you look at. What are your days of sales outstanding or DSO? These metrics are based on, are your customers paying you? You’re billing them, are they actually paying you? If they’re not paying you, maybe there’s some economic headwinds because they’re not paying you because they’re managing their cashflow more tightly. They’re stretching out their vendors a little bit, you’re one of them, and that’s going to show up in that metric.

(00:53:25):

These are just early indicators. They’re all based on trends. They all matter. You can get stuff that’s really recent and all of a sudden you see a little change. It might just be a blip in nothing, but it also might be start of something. That’s why every single day, looking at every single analysts who’s tracking stocks or if you’re an investor, you’re tuned in to every one of those metrics. You’re looking for the canary in the coal mine to try to be predictive.

(00:53:51):

That’s where our world is. That’s why the data, the speed of information, it’s so key, and that’s what gets missed because when you look at finance and even these TMS solutions, they’re all legacy. They can’t do stuff on the fly. You can’t really get into a lot of the predictive nature because the architecture or even the data that you have at your disposal is never going to allow you to play offense in that way. If you can’t play offense in today’s day and age and you’re caught flatfooted constantly, it’s not a fun place to be.

Joseph Drambarean (00:54:20):

Yeah, not at all. I genuinely, and it’s really rare that I come across things that turn me into a kid in the way that when you look at it and you think unlimited possibilities, and this really is that, and I know it’s probably hyperbole to say that this is the first of its kind in terms of generative AI for finance, but it has to be one of the first because of the approach we’ve taken.

(00:54:49):

To summarize it again, that importance of accuracy, the importance of context, the importance of privacy, those have to be bedrock in any approach that you’re going to take with regards to finance, corporate finance especially, but you go up and down that gamut, small business, mid-market, large corporates, mega corporates, they’re all going to care about the same thing. They’re not going to approach this technology if there’s a fear of any kind of inaccuracy or of lack of privacy or whatever it might be.

(00:55:19):

I think that at the end of the day, because of how we see this going, this is how you give generative AI legs. It’s taking that hybrid approach of it is a genius, but it’s a genius that is only made better if it has the right tools. Playing in that way I think will change the game for finance, and it’s not just finance. I think that this is a repeatable insight that could play out over many different industries.

Brett Turner (00:55:49):

Well, you think even the term genius, it also is a pejorative in our society too. How many times you say, “Oh, way to go, genius”? It reminds me a little bit of, yes, it’s a genius, but it’s also going to get a little bit of this way to go genius if it’s not married with the right things.

Joseph Drambarean (00:56:07):

Exactly.

Brett Turner (00:56:10):

Also, I was hearing earlier too another podcast on AO, someone was talking about just whoever has really the private databases that ChatGPT doesn’t have access to, whether it’s Quora or Reddit or things like that, LinkedIn, whatever, it doesn’t have access to those kind of things. It’s never really going to get access. We’ll see how that plays out down the road, but those applications are probably going to release versions to leverage it for their community in a way that they can do that.

(00:56:43):

We’re now the emerging leader in our field of really transforming bank data, normalizing all that in the corporate banking world or any business that’s essentially larger than a few million in revenue in on up in terms of how we do that with the platform. All the things that we do is along the lines of automation. So when you look at that, that’s all private, that’s all for the benefit as we advocate and build these tools for our customers.

Joseph Drambarean (00:57:12):

Right. Exactly.

Brett Turner (00:57:13):

So when you basically take this on something that we have that’s private and protected, but also further can be weaponized in a way with something like this, it really is profound. I mean, I think in general, this is why you’re not seeing any announcements of saying there’s any breakthroughs on AI along these lines because nobody has … They haven’t curated, they haven’t pulled together the right data set, the right tools, put all these things together in a platform yet in this way. This is why it’s going to be really, really fun as we’re moving forward.

Joseph Drambarean (00:57:49):

That’s the thing. I think the thing to look out for when announcements do start to happen are, are any of those three principles violated? Because of course, what we’re showing here is production grade. It’s something that is in the hands of customers and customers can trust it today, not next year, not this summer. It’s now. I think that that’s going to be the important thing to look for. What sacrifices need to be made by anyone else that approaches this problem space? Are there reasons? Is it because they can’t actually provide the analytics capabilities? Is it because they can’t actually interface with ChatGPT in this way? All of that will be something worth keeping an eye on because at the end of the day, what did Open AI do? They opened the floodgates to everyone using it so that they could suck all of that context and data into their platform.

(00:58:46):

It was not intended to be a private platform. It is using all of that data to further train and create GPT4, GPT5, all of the next iterations of it. So it’s really important then from an industrial use case, a corporate use case to look out for privacy because at the end of the day, you can’t really trust a large language model if you know that the very same finances that you’re trying to calculate on maybe from some other company or some other company over here, will it hallucinate? Will it accidentally cross pollinate? Those are the things that we want to make sure that could never happen in our platform. I think that it’s an important part of our approach.

Brett Turner (00:59:26):

I think the long-term approach that we’ve done is right because we focused on that big data layer at our core. We focused on security at our core. We just knew we had to do stuff. That stuff is table stakes. We couldn’t get that wrong. We had to develop that early and we had to be able to breathe confidence. That’s why we have the customers that we have because they have been able to stress test it and they saw our compliance. All those kind of things have been buttoned up for a long, long time. I think being able to now start to build off of that core framework is getting a lot more fun.

Joseph Drambarean (01:00:05):

Absolutely. So this was Trovata AI and this is not-

Brett Turner (01:00:11):

It’s going to be fun saying that.

Joseph Drambarean (01:00:12):

Yeah. You’re going to hear about Trovata AI, but yeah, this has been so exciting to work through this together with Francisco and others and just bring this to life. We’re just so genuinely excited for this to be in the hands of customers and to get to interact with something that is so breakthrough, so novel in its approach. I can’t wait to see where this goes because it’s unpredictable. Just like when Open AI released ChatGPT 3.5, and we just started to see unbelievable use cases.

(01:00:42):

Imagine that, that in the hands of customers they can ask it anything. We’re going to be able to see how does that develop over time, how does the trust go up in terms of what you can ask it and all the different use cases, and we’re going to be there, and keeping up with that pace and finding new ways to provide context and all of that. So I’m just so excited. I think this is an unbelievable development in our history as a company, and I can’t wait to see what happens.

Brett Turner (01:01:11):

Super exciting. I mean, now with almost 200 customers, you look at all the users now, two-thirds of our customers are treasury teams. The other third are finance teams, accounting teams. We have some smaller companies. We have some really big companies. We’ve got this really great contrasting and context of just all the different use cases and users and segments. It’s going to be fun to just see that collective thought and expertise to just iterate on what’s being teed up on their own data in their own private environments. It’s just going to be just too cool.

Joseph Drambarean (01:01:47):

All right. Well, this has been Fintech Corner. Really excited about our announcement and we’ll see you guys next time.