Name: Crawl, Walk, Run: Using AI in Accounting
Uploaded: 2025-05-06T19:30:10.312Z
Duration: 1 h 3 s
Description: Crawl, Walk, Run: Using AI in Accounting

Transcript for "Crawl, Walk, Run: Using AI in Accounting": Alright. Welcome. Welcome, everyone. Welcome. Welcome to today's webinar. Let's see. Looks like we already have almost two almost 300 people already in the room, which is awesome. Everybody, hello. Hello. Chat will be muted for this webinar, but you can take advantage of the q and a feature in the top right of the screen. Curious where everyone is calling in from. You guys are all very punctual. I love it. Chicago, Mohammed. Mohammed from Chicago. Welcome. Welcome. See Colorado Springs, Oakland. Okay. Shout out Bay Area. I'm originally from San Francisco myself. Linda from Las Vegas. Almost 400 people in here already. Sam, Salt Lake City. Mushtari, Dallas. Love it. Another Dallas. Nadia from Dallas as well. Awesome. Awesome. Awesome. Alright. We're gonna give maybe one more minute. While so many responses. I see Boston, Nicole. Boston, all caps. Boston strong, lots of pride there. Spent a few years in Boston myself. Perfect. Alright. Well, we'll go ahead and get started. Already over four four hundred people in here. I will let everyone else trickle in. But just wanna say, hi, everyone. Welcome to today's webinar. My name is Matthew. I create content for Rippling, and I'm super pumped to have you all join today's crawl, walk, run webinar on AI and accounting. Now, again, a quick housekeeping item. Chat will be muted, but you can take advantage of the q and a feature on top right of your screen to ask any and all questions. Do not be shy. I see a lot of you already doing that, which is great. And today, of course, we're airing an interview with the amazing Sowmya Ranganathan, former controller at OpenAI, and also formerly, at Rippling as well. Now before we jump into that, I wanted to quickly share a bit more about Rippling. Just because, you know, as people trickle in, you may know Rippling as an HR company, but we actually have a software solution for not just HR, but HR, IT, payroll, and spend management. Now the solution is probably most pertinent for today's crowd is Rippling Spend, which itself is actually an all in one spend solution. What that means is, you see, Rippling Spend consolidates your expense management, corporate cards, bill pay, and soon travel and procurement into one unified platform with powerful automations and, of course, seamless accounting integrations. Now what really makes Rippling stand out is what follows. Some of you are probably familiar with the diagram on the left, you know, disconnected, siloed, isolated HR payroll and finance tools that can't really talk to each other. They have unique data formats, very brittle integrations that break as you scale. And just by contrast, on the right side, Rippling is built around your employee data. It's always on, it's real time, and most importantly, it's interconnected across HR, IT, payroll, and finance because it's all one platform. This leads to a ton of unique use cases that save you as a CFO, controller, or manager a lot of time and money. For example, auto locking employee cards when they're missing more than x number of receipts submissions. This is actually really popular use case amongst controllers using Rippling Spend. Or blocking out a policy spend in real time before transactions can be completed, or even using AI, which is today's hot topic to turn receipt photos into ready to submit expenses. So that's Rippling Spend. Thank you for allowing the spiel. And if you're interested at the end, we'll have a gift for anyone who'd like to learn more about Rippling Spend as well. But you're all here to learn about a different way to save time, which is AI and accounting. So without further ado, it looks like we've reached, a good amount of people in here. 660 people. Let's dive in. I'm excited to share this exclusive interview with Sowmya Ranganathan, former controller of OpenAI. Alright. So thanks so much, Sowmya, for joining us today. Really, really excited to chat about using AI and accounting with you, especially given your background and your expertise in that specific space. I guess I can kick it off here, do do a quick introduction on your behalf, and then you can follow-up with another one, or any other details I might leave out here. But let's just say you you start an audit and you move from there to Square, where you kind of help them go through the IPO process. And then you jumped over to Rippling and was, Rippling's first controller, which we all still very much appreciate. And then from there, you led all aspects of controllership, including revenue accounting, opex accounting, payroll, finance data, and accounting for compute and infrastructure, which is a lot, all at a company called OpenAI. But with that, yeah, feel free to introduce yourself as well. I'm really excited to have you on the show. I know. You kinda covered it all. Thanks for having me. Really excited to chat. This is, you you're gonna have to keep me on track because I I love talking about this, and we might go in all sorts of directions. But really excited to be here. Thanks for having me. I guess before jumping into AI and accounting, we do have to talk at least a little bit, I feel like, about that moment where OpenAI launches ChatGPT. It becomes the most popular consumer app in history. And I believe your first month end close with the company was after they had just launched ChatGPT Plus. And, you know, it was the first hit, fastest to ever hit a hundred million monthly active users. And now it's some reports are saying 400,000,000, weekly active users. So I guess what pulled you to OpenAI? What had you been following the company for a while? Did you anticipate it was gonna be such a game changer? You know, I'd actually heard about OpenAI from the moment they launched, but I was never, like, actually looking for a new job or, like, following the company with an eye to, like, hey. I might work there one day kind of thing. And, I was at Rippling at the time really, you know, happy engaged. And then I at the end of twenty twenty three, I left Rippling and I took a personal spackle. I I took some time to travel, but the really interesting timing is the last day of my job at Rippling was the day chat to BT launched. Oh, wow. So the whole time, like, I I I was on a break and I wasn't working, but so I had a lot of time. And I was just I couldn't stop tinkering and playing around. I was like, oh my god. How is it so smart? It's doing all these things. And, you know, I was like, oh my god. Like, what if I could build a little, like, you know, accounting GPT where it answers, like, technical accounting questions? What if I could, you know, put this on top of NetSuite data? And now you could finally, like, get data out of NetSuite without figuring out all the clicks. And I I was kind of in this really enamored state of, like, jaw dropping, like, what is this thing kind of technology. Mhmm. So after I came back from my trip, I was back in San Francisco, and I saw an email from a recruiter at OpenAI. But I think the thing that pulled me in was, one, this kind of realization of they have the same struggles. They as in, like, the finance and accounting team at OpenAI. They have the same struggles as any other start up I've seen where the company is doing really well. You haven't, you know, built enough kind of accounting infrastructure. And so I thought, wow. Maybe I could actually bring some of that, like, scaling experience to OpenAI. By the way, the growth at OpenAI humbled me very quickly. Like, it was like nothing that's seen before. It's like, okay. So I guess we're figuring things out from the ground up. Yeah. It was it was kinda to the point of, you know, you would launch something today, and you you have about a week to figure out how you're gonna do this at scale because there's no grace period for doing things manually. Like, basically, if you can't automate it in, like, two weeks of something, like, new happening, you can't close the books that month. And it it was, like, nothing I'd ever kind of experienced. I usually, you know, even at the fastest growing companies, like, Square was going really fast, Rippling grew really fast. You have, like, about a month or two at least where you'd kind of get your footing and be like, alright. What am I gonna do? Here, it was kinda like, go. That's it. Yeah. I mean, when you have such a viral product that's doing so well, the second you lost something, probably millions of people using it within hours. So I could totally see how that happens, and I can also see how that kind of I think you've mentioned this before, but it forced you to figure out how to automate a lot of these things and and use AI, to speed up your month end close and just, smooth out every part of your process. But that's kind of a serendipitous story. That's really interesting. You know, your last day of Rippling, it comes out, and you had so much time to play around with it. And, yeah, that's awesome. That's awesome. I think it's rare to be able to see something so game changing and then be able to go act on it and be involved. So that's fantastic. No. I feel really lucky, honestly. And so, I mean, did you you must have somewhat anticipated, like, this kind of growth curve, but maybe, like you said, to not the extent that it turned out to be. But how did you prepare your system and your team in the accounting function for what would be, you know, the so many launches at at OpenAI? I think, to be honest, I was just kinda really obsessing about the product where I wasn't even expecting like, I I didn't have any kind of what's the business gonna do, and the business kind of blew all of our expectations. Mhmm. I think the one thing that I'll say is the company has already, like, a very intensely talented team in place when I joined. It was a very small team. Like, I think accounting was around five or 10 people, all including, you know, operations, everything, but really, really high talent bar, and they'd put in, like, reporting pipelines for revenue already for the existing revenue. Right? Like, we already had API revenue even before Chattoport revenue. And I think the notion that you already had revenue running on SQL and not Excel, as simple as that sounds, was kind of the make it versus break it moment for the team at that time. Even if it was painful and, you know, we, of course, like, didn't have dedicated engineering time, engineers were just trying to keep the site alive, We still had some kind of, like, a a really good foundation actually to, like, build and run revenue at that scale. I I look back to then versus now, and we're, of course, like, much better set up to run revenue where it's not kind of a fire drill every month. But if if we didn't have the peep those people and we didn't have the things that they've built for, you know, say, the year of leading up to that, it would have been just, like, I don't know. I might have just, like, walked in and walked out and be like, I I don't think I can do this. So a really big shout out to the team. And, you know, it it kinda tells you, like, even if you have the best tech in the business, you kinda need the people that can, like, be creative and problem solve leveraging those tools. Totally. I think outside of revenue actually was where we ran into most of our problems. Because if you think about it, we ran revenue through Stripe, and Stripe gives you this neatly packaged data and this data pipelines, and you can do things with it. It's structured. Like, it's it's actually a really good foundation to build on. But for OPEX and compute data, it's like you you have to go to some vendor website and download a CSV of activity. And, like, how are you gonna do those things in SQL or there's no existing system, there's no tooling. It it felt much harder to solve problems at scale. And by the way, this is not hypothetical. This happened Because when your revenue grows, your compute cost grow, you know, you're you're basically melting those GPUs. And the activity file for that blew up where we couldn't open the month of activity in Excel anymore because there's too many rows of data. Right. So that's kinda what I mean when you say, like, there's no grace period. Like, it just breaks. And then, like, you you're you're staring at, like, a very broken thing. And that's that's what work felt like for a long time. Yeah. It definitely keeps you on your toes. And I think there's a lot of good insights just in what you just shared. And I think what we'll try and tease out across the rest of this conversation as well is, you know, being at OpenAI, you eventually had to build this process that would scale and would be able to, you know, adapt within the matter of couple of weeks. And I guess at a higher level, and we can dive in deeper as as time goes on, what do you think were the key differences or contrast between the system and, the processes that you built to support this unique hypergrowth at OpenAI versus, let's say, a traditional accounting accounting function that you'd stumble upon in the average company? For example, like, having things both on SQL instead of Excel by the time you got there. Yeah. I think if I had to generalize and abstract this, I think the biggest difference is how we thought of finance data. I famously traded a revenue accounting headcount in my first month there for a finance data, like, engineer like, very technical, full stack data engineer role. Mhmm. And it was you know, in in most companies I've been, finance data is seen as kind of this downstream reporting layer. Like, once accounting and, you know, the finance team does their job, like, you you do dashboarding, metrics, reporting, and it's very, like, downstream to where the work happens in many ways. We kind of turned that on its head completely, and we put finance data in the middle of everything we did, which means that this person is embedded into our finance team. They sit right next to like, you know, that this we had one guy, by the way. I'm calling it a team. Now it's a team of two. But for the longest time, it was one person. And he would sit right next to the revenue team, and I said, you know, your your mandate is to make revenue, like, touchless. And what is it gonna take to do that? And you you kinda realize, like, if you put finance, like, you know, you you have a data warehouse. So at the time, I think we had like, we're primarily using Databricks at OpenAI. And you essentially have data from your upstream systems flowing into Databricks. You you can do, like, your transformations, queries, Python scripts, whatever you need, and then you export data out of Databricks into NetSuite. And then NetSuite data comes back to Databricks, and you can run recons between your source data and NetSuite data. So I think having this, like, kind of you you put your data warehouse in the center, and then you have all the pipelining and integrations up and down and whatever else gives people the ability to then use Chargebee T or whatever else to, like, you know, you you it can write SQL for you, but where are you gonna run that SQL if you don't have a data platform to do that on? Mhmm. And so I think the the foundational investments we made in finance data was the game changer. And, honestly, we did this even before other parts of the company. Like, other teams saw what we did, and they're like, I want a data team too. That's awesome. That's awesome. Yeah. I think that theme is gonna come up a lot today. So, yeah, that that's some really great stuff right there. So thank you for sharing that. Yeah. So I guess we can start to dive into, you know, how do we actually use AI. I know you have a couple of, walk throughs prepared for us too today, which is really exciting, to be able to get so hands on. So we're kind of breaking this into, again, a crawl, walk, run framework, so that, you know, there's something for everyone at every stage in in their process of learning how to use AI. And I guess starting at the crawl stage, one thing maybe people don't realize, but is perhaps a blocker to starting to use AI is how do you create a culture at work and in your team that embraces AI rather than maybe feeling feel fearful of it? I think the main thing to realize is AI is really new technology. Right? So the only way to give, the only way to have people adopt it is to give folks a really low stakes way to explore use cases because not everything is gonna pan out immediately. Mhmm. You kinda have to find your most pressing problems. You build MVPs around it. And and and you kinda say, okay. What is kind of the most problematic thing that I do manually? And what if I could just use, you know, Trygbt to try to figure out how to automate that small bit, and then slowly expand? One thing I I really love, and it it really worked out well for us, is to host these team hack a thons. So we would have these, like, team ops sites, and I would take, like, two, three hours off of that schedule and say, this is gonna be, you know, finance hack a thon. And, like, finance teams don't usually do hack a thon, so what even is that? Turns out if you give people an opportunity to drop in their ideas for things they wanna work on and self organize into small groups, you can actually make a lot of progress just like poking around with Chagibuty for two hours, and then you do a demo for for an hour. And the minute you have, like, prizes I don't know. I'm I'm really competitive like that. And, it's it it it's just it's like it doesn't feel like work, and it feels like, you know, I'm not being measured for my performance review on this stuff. It's truly just like a fun new thing I'm exploring. I think turning it into that kind of thing has a huge impact in how people feel about adopting it. I'd say if I'd gone in and, you know, given kind of a team mandate, like, we're gonna be AI first, dah dah dah. But, like, I I think it would just add a lot of pressure and not have a lot of outcomes if if you kinda went a very top down approach. I think in the long term, we adopted kind of a you know, both things happened. Right? You we encourage bottoms up adoption, and we continue to host these hackathons. But then we also, like, have OKRs and other, like, kind of performance goals tied to what could be, you know, streamlined and out of it. Got it. Yeah. I love that approach. I could totally see how, you know, saying, we have to implement AI somehow, like, figure it out can be more stressful than, like, hey. Let's just see where our ideas and creativity takes us. Or, like, you know, sometimes people know their own pains more than, like, maybe their managers do, and they'll come up with a use case that's actually, really impactful for them, that maybe might be hidden otherwise. So, yeah, that's great. That's great. And so I guess we can dive into the first kinda entry level use case here, for ChattGPT to help people get comfortable. So this would be, you know, for accountants, something that's very pretty basic and just to get you used to interacting with ChatGPT. This is something that people, if you're watching, you can follow along with maybe. But, yeah, we can we can dive into the example and maybe even just talk about at a high level what else Chargebee team can help people with so they can kinda understand how to think about the tool. Yeah. I'd say before we jump into the tool, the thing with Chargebee team and and and, by the way, like, we're talking about Chargebee team today, but you could you know, I I I I'm I'm not biased where you you can pick pick the tool of your choice. I'd say, largely, they behave similarly, and you'll you'll start noticing, like, slight personality preferences of the AI models and such. But today, we'll be demoing ChargebeeT. I'd say you you log in and you see this big empty box. Right? Behind that has a lot of capabilities that are not immediately obvious until you try to, like, do things with it. So it's a writing assistant. It can do technical memo writing. It can, you can basically take you know, you can write your strategy planning doc. It can write basically whatever you want. But my thing is I try to treat it as a little bit of an assistant. So I I dictate a lot of the core content, and then I I keep asking it for edits until I feel good about it. And sometimes, like, once it's in in in, like, a close to final draft, I will basically, like, just go ahead and make edits myself. Mhmm. It's a good research assistant. If you haven't heard or tried Deep Research, I'll show you an example later, but it's really phenomenal. I've used Deep Research for, like, totally random things. Like, honestly, a lot of purchase decisions in my life are now Deep Research driven. I used to obsess with, like, a lot of Internet research about, like, this one category of things. Like, I bought this, like, digital like, a mirrorless camera recently, and normally, it would have taken me kind of, like, weeks on weeks of Reddit and, like, going into random, like, parts of the Internet to find out. And now, like, deep research is really good at helping you think about pros and cons and trade offs and whatnot. And you can, of course, do this for work too. In accounting, for example, you can use deep research to benchmark SEC disclosures, and you can say for this topic, what are other people, in my industry doing? These are the peers I think. Like, are there other peers that you think I should be looking at? And, like, it's really good at kind of peeling back the layers of the onion and going as deep as you ask it to. Mhmm. It, of course, is very good at coding. If you have friends who are engineers, they will tell you that, it it it's gotten really quite good, and there's a lot of auto completing. Coding is it's it's it's I say it's it's a little bit of a harder hurdle to do it your first time, but once it once you start using it, you can't go back. And you almost have to retrain your Excel muscles to say, you know, pause. Don't don't, like, build your spreadsheet yet. Try it in this other way. Mhmm. You know, I've used it for, like, formula help. Like, if you have this horrible, like, nested if condition formula and you can't keep your parenthesis straight, you can just ask it to write it for you or just fix it for you. Yeah. But but I think my favorite one is to learn new things with LGBT. Matthew, like, when we hire people for revenue at OpenAI, we wanted to hire people with revenue experience. Right? But the way we actually did revenue accounting was all, like, SQL based. And if they didn't do SQL, it would be really hard for them to be good at their job. Mhmm. And we we set out, you know, with, like, very optimistic vibes on, okay. We're gonna find a revenue person that knows SQL. And turns out that, like, Venn diagram, the intersection is so small. And, you know, like, we we we have really great people that at some point you have to make a trade off. And we said, okay. We'll hire for revenue experience, and then we will train you on SQL. And it's really phenomenal how many people are now total pros at SQL just by, like in the first three months, they basically asked you to fix all their queries. But then over time, they don't need that anymore because they've just learned by doing in a really kind of, grounds up way without needing to take, like, a SQL course or an exam or, like, anything like that. And so I think if if you really kinda get into it, you can use it as a super powered tutor to teach you, like, whatever you needed. So shall I actually, like, do a screen share? Yeah. Let's do let's do it if you're comfortable with it. That that works. Let's do it. This is my ChatGPT. I use ChatGPT for basically everything. So okay. So this is an example of actually deep research, and you can do this without say you've never used ChattGPT before, and you're like, what can this do for accounting? I uploaded an order form. Okay. This is actually a ChattGPT generated order form. So it's not a real company, but I've put in some, like, you know, there's a three month free term, and there's overage fees, and there's an implementation fee, and, you know, some such. Right? So coming back to ChattGPT, I basically uploaded this order form, and I gave it, you know, really high level instructions. Use your six zero six knowledge, use the attached guide, and document your conclusion. Now my instruction to it is really generic. Right? I haven't actually told it much. I basically like, if I gave my team something like this, they would kind of just roll their eyes at me, and they're like, what do you want? But there's no eye rolling here. It's asking internal facing? Do you want this to be audit support? And then what I do is I copy paste that, and I basically give it a response. So here I'm saying, yes. I want the five step model documented. Internal memo is fine. I don't want prelim guidance. Just give me a final position. So it took nine minutes. And the reason deep research takes a little bit longer is if you go down, you can see that everything is cited and sourced. Even if you don't need it to be that way in your own memo, it's really helpful when you're going back through it and being like, why did you think this? You can go and cross check yourself back to the source. But here is its output, and I'd say there there's nothing else I've really given it. Right? I didn't give it, any examples. I didn't give it any feedback. This is just, like, the first shot output. It's putting a background together and, honestly, goes through the five step model pretty well. Now if if your company says most of our order forms are basically the same and we don't need this, like, super full on memo for every single order form, I just need you to put it in a table format. It can do that. You can then say, okay. Now that you've documented the table in this way, extract key contract terms and put it also extracted in this table, and then you can, like, you know, keep going. Right? You can basically say automate my rev rec calculation, create my different revenue waterfall. Like, you just kinda build on it. But the point is the latest models, my favorite, honestly, is deep research, and you're gonna, like, hear me say this a lot for technical accounting, but you don't need to do much more than than just, like, ask the question and get going. So if you're wondering where do I get started, I would say without doing much more, you can start with this. In general, you know, it it might be a little counterintuitive, and I I know I just talked about how good ChargebeeT is for learning new things. But if you're trying to learn how to use Chargebee T, my recommendation is actually to use use it for something you're really familiar with, a journal entry that you do every single month, a report that you prepare every month, a presentation you know inside out, And then start talking to ChargeGPT about it, and then see if you can get it to recreate it or ask it for feedback on ways to improve. That way, like, you kind of learn how to interact with ChargeGPT on the topic without also being confused about what the hell is it trying to do. You know what I mean? Because you know the thing that you're working on really well. I would say, you know, I I can keep going, but maybe we'll flip back and, like, I'll I'll show you some other, like, cool tidbits down the road. How does that sound? Yeah. That sounds great. Yeah. I think you're walking through that example. And I think, to your point about using chat g b t was something you're super familiar with as a way to get comfortable with it, I've definitely done that before. And it helps me also trust it more because I can know immediately its outputs whether they're right or wrong. And that kind of brings me to another question around, you know, if people are worried about ChattGPT or AI making mistakes or hallucinating or they feel like they can't reasonably conduct quality control if its output is, like, thousands of rows of data or something like that. How would you recommend building that trust with AI's outputs? And, you know, how how can people go about that? Okay. So a few things. And you'll actually see maybe some of this. I have an example of, like, working with the CSV file, and you'll see I'm basically, like, cross checking it at every stage. And I think that's actually maybe one of the key things. You start small. Don't try to get it to solve your entire problem because it will come up with some solution, but you don't know if it's right or wrong. Right? So you start small and test the output at each stage. For things like technical memos, you should treat it like you would treat anything else that's been prepared. It it it's a great first draft, but if you're gonna put your name on it, you're gonna review it, you're gonna, like, add your own commentary, you're gonna make your edits. And I would say just because it came out of ChargeBT, it doesn't mean you can automatically just, like, ship it. I I know people do, and I I would highly say don't. You know, like, it it it's it's still something that you're shipping, and you wanna make sure you can put your name behind it. The nice thing, though, is it does speed things up quite a bit. Right? So you don't need to, I I would say if you have any sense of how long things should take, you should kinda throw that out the window because sometimes you'll get it done really quickly, and sometimes you're kind of like, wow. I should've just maybe not done this. And, like, the more you use it, the more you kind of build that compass for what what is a good use case and what is kind of, like, a little bit of a waste of your time. For data analysis, I I think for most finance people, if you've not seen or done, like, in any kind of programming, it it can feel a little daunting. One thing we'll say is Python code, there's a lot of syntax when you want to write it, but when you're reading it, it reads like math and English. So you can actually read it top to bottom, and it it it it it actually will make logical sense. And so don't feel like, oh my god. It spit out this code, and now I have to run it, and I don't know what it's doing. I would say kinda just, like, work through that initial barrier and still read through. Sometimes it will come up with variable names and, like, you know, just just think of it as that. And, like, you can call it x and y if that that is your comfort zone, but, it it's it's really worth that time investment to still lead the logic. And, of course, test it test the script on historical data. If you've used the same process, like, for many previous months or years, like, put that old data and see how it compares and actually spend the time to reconcile the differences. Funny enough, when I started doing this, I found a bunch of bugs in my manual work, and it was kind of a it was not a happy moment for me where I was like, goddamn it. Like, where were you two years ago? That makes it a little hard to quality control when you're not sure which ones. Yeah. I mean, you know, you you still have to reconcile just like you would reconcile any kind of new methodology. But I would say just because it's AI, you know, all all of the practices we know from, like, you know, all the kind of analytical work that you people do over the years. Like, you you kind of bring those same practices to AI as well. Definitely. Yeah. I think, I'm definitely hearing a theme of, like, just spending more time using it, starting small, like, allowing it to teach you, and you also learning how to use it properly. So with that said, with everything in mind, we can kind of jump to the walk section and talk about the next level of these AI use cases for accounting. So some of the ones that you've mentioned before that I think are definitely really useful, are, you know, for example, processing huge CSV files instead of fumbling with really laggy Excel sheets or generating reports and reconciliations with SQL, joining data from multiple systems for simpler reconciliations and reporting, and forecasting and modeling with Python. Even for advanced users, bring in some analytics and machine learning teams for really in-depth scenario analysis. So I know we probably can't demonstrate all of these within an hour even though I know you probably want to and everyone wants wants to hear about them. But maybe you could give us, like, a high level overview of how you would approach these with AI, and then we can dive into maybe, like, the CSV, example for everyone today. Yeah. I'd say one thing that might be helpful for people is to say, how does this actually fit in into your closed process? You know, you're gonna show me this Python code, but does this kind of live in a silo? So coming back to kind of the initial example I shared where as the product grew like crazy, our usage grew like crazy, our compute cost file went kind of ridiculous where we had to, you know, break the month of activity into nine or 10 different files and then do, like, this Excel pivoting, vlookup, whatever, nine times and then, like, stitch it together. And it was brutal and still broke people's computers. And it was it was just, like, this massively painful thing. Now fast forward, to now, the the usage data from Microsoft and OpenEdge case went straight into Databricks through, like, a blob storage integration. I'm gonna be dropping some names. Don't feel like you have to understand every single, like, terminology. It'll be different for your own company based on, like, what data architecture you have. But, essentially, think of this as, some kind of, I don't know, SFTP transfer from your vendor into your data warehouse. Right? And from there, you can set this Python code to run on like, it it it automatically updates the dataset, and this code is running automatically without you needing to do anything. And instead of, like, numbers that you put together at the end of the month after close, any stakeholder can basically go to this dashboard and real time see how your costs are trending. So in our case Uh-huh. One, we wanted to see, like, what is the cost to date, right, and how does this track against, budget. But we also wanted to see how does how do the deliveries tracked against plan? And, like, are we ahead of plan? Are we behind plan? And these are super material to our business, and I think just the benefit of having this in real time is huge. Just, like, think about knowing what's actually happening I mean, happening on the 15 the month versus waiting until, you know, ten days after the close of the month to, like, try to cobble it together. By that time, you feel like, oh my god. I don't even like, we're we're in a whole different place in the business where the numbers you show me don't make sense anymore because it's just not relevant. Mhmm. And and from that automated dashboard, this goes back to what I said about putting data in the center. You can actually build pipelines, reverse ETL pipelines back into NetSuite to then just automate your accrual j for that month. So you can kind of have this entire process be touchless, and it it it it is a really good dream state, to be honest. Like, it's something that I've been wanting for so long, and it was really magical to kinda see it come to life. But the starting point was what we're gonna do today, which is literally just dump a CSV into chat g p t and start, like, changing how you do it. Today's example, of course, I'm gonna oversimplify a bit. It is, you know, you're gonna look at it and say, I could do this in two seconds in Excel, but that's not the point. You can, of course, take much more complicated logic and turn it into Python. So without more preamble, let me show you what I'm talking about. Perfect. And I think you mentioned, ETL pipelines as well, just for anyone who's not familiar. What does that stand for? Extract, transform, load. Yeah. When I say ETL, it basically is like the raw data from your provider might not be in the format you want. So the data people in in our company essentially help they'll they'll be like, okay. What do you want this table to represent? And then they can stitch together, or or do kind of transformations of, like, they can apply business logic to make sure that the foundational tables you're using, make sense for what you need it for. Got it. Okay. So coming back to this data analysis example. So I've done this example yesterday in the interest of time, so I'll kinda walk through the chat. You'll notice that I'm starting with just the column headers. Right? And I just said, I have a CSV file with these columns. Now if if you're worried about sensitive data and or if your data is just too big and you can't be bothered to, like, upload the whole file or type GPT says your file is too big, the the cool thing with this kind of, approach is you don't actually care about the data at all. You only care about the headers because the things you're doing are at the column level. So I start with the headers and okay. It understood what this is, and it's like, what do you want me to do? Now this is how I talk to ChachiPT, which is a little hard to read for a human, but you can basically just go stream of consciousness on it. Right? You can say write me a Python script, filter out this value. Right? If anything has subscription one, take it out. I don't care. I want you to do some mapping. So if it's subscription two, put it in cost of revenue. Three and four are r and d. Five is s and m. If there's some other subscription value that I don't know about and I haven't explicitly told you, just dump it in g and a. This this I like doing this because sometimes if you don't put a catch all in there in the logic, it'll only do the things you told it, and then your totals won't reconcile. And then you're kinda like, what is this? You can also call g and a into, like, a different thing. Like, you can say, you know, unmapped or something. So you know when you see that in the output that you're supposed to go back and change your mapping logic. And then I'm saying, okay. Once you have this mapping, you know, sum the values based on this mapping, and then what should the final output look like? If if you put this in an Excel world, I'm basically doing some filtering. I'm setting up a mapping table, and then I'm pivoting it by the map. So it's like this is what I meant when you when you if you're watching this and thinking you can do this in Excel, just bear with me and, like, kinda see what's happening in Python, and then you'll you'll the way you describe the logic is basically here. You can ask it to do whatever you want. And then it, of course, is very quick at spinning out this Python code. And it's this is what I mean. It's saying import this is all English. It is this is some some kind of, like, okay. It takes your CSV and it puts it in this thing called a data frame, which it's calling d f for short. And then it's, like, doing your mapping here. So when you want to cross check what it's doing, you can basically double check to say, is this what I told it, or is it, like, doing the wrong thing? So I encourage everybody that is doing this to still, like, spend the time to read it. It might not all make sense. If you if you really care, you can just copy paste that back into chat GPT, like, what exactly are you doing in this line, and it'll tell you. And then it's saying the output will be a table like this. Okay. Pretty good. Now I'm saying I actually it it won't show you here, but I uploaded my actual CSV file, And then I said, run this code on my sample. Right? This is me doing back testing essentially for, like, a subset of data where I can do this in Excel pretty quickly and compare the output. So look what it did. It put everything in g and a. And immediately, I was like, this isn't right. Why are there no values? So this is what I mean. Sometimes, like, it's really good, but it's also kind of it it'll do some really ridiculously bad things, and you always have to sanity check that out. But just like you would when, you know, somebody else, like, a not an AI bot does it for you. And now it says, well, good cash. It means it doesn't match exactly because I said subscription two without the underscore. Uh-huh. And but I didn't even have to teach it how to fix it. It it it was just like, oh, it's because you said it without the with the dah dah dah, but, like, actually, it's like this. Let me fix the mapping. And in theory, I should have caught it here. Mhmm. But I was just kind of skimming over. Right? But, like, part of it is don't accept the first output, but when you tell it something's wrong, it's, like, much better at going and fixing it. So now it says, okay. This looks much more like what I'm expecting. And then I said, okay. Format this because, like, you know, sometimes the numbers get big and I'm like, I don't know how to read this. So I can, of course, use little things like that. Now I'm, like, saying some of the total values by each ID so I can make sure it's, like, you know, actually mapping and applying it in the right way. And then it did it, but then it dropped subscription one. Right? And I was like, well, how do I know you're filtering out the right amount, when when I'm adding it back? And now it complains, and it says, you told me to filter out subscription one. I was like, alright. That's so I I don't know why I read it with SAS, but, I just imagined it. And then I said, okay. Fine. Just like I know I said it, but just like put the full table without exclusion and then it does it. It's really funny. And and now you can basically see, okay, this value is gone in my output table. This one, fifty four three one one, is this one. And then I'm pretty like, I'm not great at mental math, but this is, like, you know, 60 plus 30 something, like, 93 ish. Okay. There you go. So now I can see pretty quickly that what I had in my dataset is what it's doing. So Mhmm. You know how we talk in abstract terms about how do you test and how do you check the output? I wanted to give you kind of the full experience of this conversation because it's not it's very rarely that the first thing you'd get is the perfect thing. And sometimes it's great and, like, the models are getting better, so it's more likely to be that way. But in my experience, it's been much more of, actually, I meant this other thing or, like, you know, you you there's this kind of back and forth, and just don't shy away from it. Nobody, in this case, you guys were all watching the way I talked to Chatt GPT, and I'm a little conscious now. But in most cases, nobody really sees what you say or do and, like, you know, there really are no dumb questions. So it it's just a matter of trying to think about your own, you know, spreadsheet logic and more abstract terms and turning it into if you weren't teaching somebody how to do it, how would you do it? Mhmm. Yeah. No. That was super helpful. I I love seeing the raw exchange of, like, how it actually goes down. You know? I think that's a lot more beneficial for anyone starting to use HBT than, like, a perfect demo. And, yeah, it just shows that back and forth that you need to kind of learn and get comfortable with, and, you know, eventually, you'll get there. But I I appreciate it. I enjoyed your narration. So, the next question I had was kind of around, like, how to centralize your data, like, the tech stack you need. But I guess as we've had this conversation, I'm realizing maybe what's more beneficial would be, like, any other another example of a use case for someone who's getting started rather than jumping ahead to saying, you know, here's the tools you'll need if you wanna completely revamp your transformation over the course of, like, the next year or two. So I think you had mentioned maybe there is, like, an agent that you had built as a quick little demo for today as well. So, would you would you wanna jump into that? Yeah. Let me show you really quick. So the the premise of this is imagine, you have a journal entry you need to make every month for sales tax, and you have all of your transactional data in Stripe. If you're a company like OpenAI, you have a, you know, a large number of, like, smallish dollar subscriptions, and they add up, you know, in aggregate, but it just kind of blows up your transaction file. So what if you, could write a bot that just says, look up all these, transactions, categorize them by US or non US based on what the jurisdiction name is, and I want you to put US in one liability account, non US in a different liability account. Pretty simple, except when you get into, like, codifying this. Imagine country code wasn't a column, just, like, in theory. You would have to, like, like, how would you write that Excel logic for US versus non US? Like, you have to essentially teach it some kind of, you know, two letter acronym kind of mapping table, and you have to maintain it. New countries are getting added, then you have to maintain it. But with ChargebeeT, because it is an AI model, you don't have to even worry about that. It just knows, like, US versus non US automatically. Mhmm. So I picked this as an example of things that you might have to do. Things like data cleanup is an example where in Excel formulas, you need to be super precise, but the fuzziness of your logic is much more friendly in a chat GBT world. Okay. So let me show you is it helpful to see how you set up this to begin with? Yeah. Really helpful. Yeah. So I have a sample file here. Clicking it just downloaded it. The I wanted it to preview, but it's okay. I'll just say, you can imagine this is my monthly, like, sales tax, you know, dump from my provider. And then I'm gonna say, can you please run the GPT and create my JE? Let's see if this works. Okay. Here's the preview. So it has, like, a bunch of invoices, transaction date, amount, tax code, jurisdiction name. There is a country code here. And so if Chattopity was smart, it would just use this for US versus not US. But even in cases where, it is not there, it can basically use something like this to pretty easily do that mapping for you. And then it has a bunch of different columns. Keep in mind, all I did was upload this and then say run my JE. Right? And how does it know to do all this? So first, it's doing top eight jurisdiction by tax amount. This could be helpful for your tax team to try to figure out, in do I need to go register in these countries? Is there am am I tripping some threshold? And then it's actually spinning out a journal entry. It is putting a debit to AR, credit to domestic, credit to foreign, and it's doing this. Obviously, my numbers are pretty small here because it's my example file. But what if you could teach any recurring j that you do to chat g p t and you're just dumping in the new month of data and it spits out a general entry? So the way you do this is you go to edit g p t. And here's kinda what I've done. It's also just plain English. I am basically saying you're gonna get a raw monthly file. Use the data analysis tools that you have. Summarize this tax amount. Present this in a chart. Like, all the things you saw happen are here. And here, I'm basically teaching it the journal entry. I'm saying this is exactly how you should, you know, categorize them, use this GL account for this, this other GL account for that. The total debits should match your credits. And all of these instructions, you put in once, and that's it. It just remembers it. And that way, like, you can do this for your equity jays, any kind of, you know, monthly activity that you have that has the same input. You could actually also do this for, your closed reporting. Right? You could say, I'm gonna upload my month end financial. Like, this is my monthly PNL actual. This is my PNL budget. I want you to build these charts for me that you can then put in your monthly close down. There's kinda no limit to this. You can train it to do anything. The CFO at OpenAI, we would give her this, like, 80 page closed deck every month, and she had her own GPT to summarize, like, the top three or five insights. Mhmm. And it was actually really good. And, like, it can look at it it gathers insights that are not even explicitly stated as a bullet on the slide. It essentially is reasoning through the things and, like, finding patterns that, you know, you might not have explicitly been called up. So that's a GPT. A GPT is kind of like you know, think of it like a try GPT template. There's a lot of buzz around agents. To me, this is not quite an agent, but, like, it's kind of the beginning of it. And a real agent, in my opinion, is if it goes and grabs this data automatically post it in NetSuite, then, like, I'll give you agent title. Yeah. That'd be, like, crawl, walk, run, and sprint maybe. You know? Yeah. Marathon. Yeah. Exactly. And I know we're coming up on a little bit of time here. How do you actually create, the prompt for that GPT? Like, how did you test that, separately, like, in a separate conversation, or is there a way people should think about creating their GPTs and the the prompts there? So, if it's helpful, I can share again really quick. But you can actually test while creating it. So there there's actually two ways to create a GPT. You can go here where you just type anything quickly, and then ChatGPT cleans up your instructions and puts it in a nice way for itself. Like, that's actually what I do. Like, I didn't know the right proceed to use code interpreter to analyze the data. TagPVTT added that. Mhmm. I just said, I'm gonna give you this raw file, dah dah dah. I don't have my history of when I created this right now. Sure. But you see this, like, box on the right? You can essentially test with sample data and see what it's gonna do. So it it's very, very similar experience to just, like, using ChargebeeT in general, except your you can do this for yourself. You can do this for your team. We have, like, an AP coding GPT at OpenAI that someone on my team created where you can upload any kind of invoice or say, hey. I have a new vendor. You know, what should I do? And the mapping is already there. So you you can imagine how you can use this for, like, you have new people on the team, you can onboard them a lot easier. Mhmm. And you save yourself a lot of Slack things. Got it. Yeah. That's that's always nice. Yeah. Perfect. Well, I had a couple more questions. For the run portion, I want to think about the future and, not just AI, but how actual finance and accounting functions will change and adapt to, be able to use AI in an even better and more effective way. So one is, how do you think an accounting or finance team structure changes five years from now, once everyone has kind of these these AI capabilities and and they're able to really use it fluently. And in that sense, which accounting or even non accounting teams do you think will see merge or even disappear? Yeah. I think, honestly, the reality of where AI is today is the teams of the future are gonna be smaller than teams that we've seen in the past. Right? And like I said in the very top of the call, the tech is really good, and you still need really strong kind of business oriented people to really leverage it and have those creative solutions. I think, the teams of the future will be smaller. They will have more breadth. They will be more plugged into the product and the business and, like, what are you trying to do? You know, there's a lot of kind of talk about agency and what people really need in this new world is agency. And what what I mean is, like, you see a problem in front of you, you need to, like, recognize it's a problem by yourself and, like, go try to solve it by yourself. And, like, you'll have incredible tooling available to you to go do that, but no one's necessarily gonna come and tell you that, hey. This is your job now to solve it, because the solving itself happens, like, somewhat quickly, if that makes sense. It's It's a little bit abstract, way to think about it. But I do think the people that know how to leverage these tools and make their work, you know if if you take a lot of pride in your man of work, right, I think you need to kind of, one, slow down a bit and really question, like, why is it that you're taking pride in that? And, you know, you you have to channel that energy and kind of pride in work to a different way of getting things done. Because over time, you know, if if you were the founder of the business, you care about the outcomes. You don't really care about, like, how long it took someone to do something. Right? And if you if you get into that mindset, you realize, like, you're not necessarily gonna get, you know, extra bonus points for taking longer. In fact, people that know how to get things done faster and more efficiently are probably gonna, like, lead the pack. Mhmm. Yeah. I mean, that that leads right into the other question I was gonna ask, which is, you know, what what are these roles or skill sets that'll be absolutely standard, you know, five years from now? And I think you definitely touched on that there, was improving your fluency and using tools like this. And I think today was a really great starting point and jumping off point for a lot of people. And the last thing I wanna ask you about is is actually, your new venture, the the what you've been working on since leaving OpenAI. You started Lumera, which is helping finance teams deploy AI that actually works. So, getting really tactical and and and showing them how they can improve their processes. So, yeah, could you could you tell us a little bit more about what you're hoping to do and and and what are you doing day to day now? Yeah. I think one of my big reasons for I'm honestly pretty crazy to leave OpenAI right now. It's it's a really fantastic time to go do that. I think, part part of it was I've had this itch to try to, like, build something out on my own, and I wanted to give myself the time and space to do it. And I think Lomaira is a really good kinda jumping off project for that. I I am helping people in a few different ways now. Right? One is I can do what we did today, which is I will listen to your bottlenecks and, you know, the most pressing problems, the things that are making you unable to finish your close on time, the things that make your reports too stale and you wanna turn it into a real time dashboard. And I I could my goal is to bring really practical applications really quickly. You know, there's a lot of kind of system implementations that take eighteen months in the world of finance. And I think with AI, that world is changing really quickly, and you can get a lot done very fast. So it's it's a little bit of deep dive advisory that's very tailored to companies. The other thing I'm focused on is really bringing that enablement. I felt like you know, I've been going on all these things and talking about how the people that know how to use AI are gonna do really well. And I thought it's kind of unfair if you don't also, like, provide those opportunities and platforms for people to get that hands on, experience. So I am doing, like, small group training sessions, and I'm always happy to chat about this. If If you have a project that you're working on and you want just a little bit of unblocking or even just, like, moral support so you don't give up in the middle, like, I'm I'm here for all of it. And lastly, I want to put a little bit more content out there for to hear from practitioners, on, you know, what's working, what's not. Matthew, one thing we didn't actually touch on is how do you think about plugging AI into your team. Right? And I think my my philosophy on this is kind of two part. One, for all the systems that handle your transactions, pick the AI enabled system because it automatically is more efficient. So Rippling Spend is an example, right, where it it has AI embedded where the receipt matching and, like, coding happens. That's already helped by AI. This is true for, like, your revenue contract review. This is true for, you know, your month end close checklist. It's kind of across the board and not you know, you know, if you have a high volume transactional part of your team, you want AI to be embedded in there so they're not constantly, like, uploading, downloading from ChargebeePT. So when you're thinking about what your tech stack is, that's something to really keep in mind. It also kinda tells you go with the vendors who are being, like, innovative and ready to kind of go with the flow because this tech is changing really fast and you wanna bet on the vendors that will, like, be at the forefront of it and therefore, like, kind of pull you along kind of automatically. And the other part is give people in your team access to these tools because, you know, nobody could have, like you know, the person closest to these horrible files and things that break are the ones that are best suited to go and automate them. And they might not even describe it to their friend. They just kind of grapple with it by themselves on the computer. And just having this assistant or Copilot or whatever, like, right there and the the barrier to start using it should be so low that people just automatically use it. So I think there's room for both of those, and I would say they you should think of them as something that happens in tandem and not so much as a replacement for each other. Got it. Yeah. I think that that point about embedding AI into your team, not just through your own interactions with something like ChatGBT, but in the tools that you use is a is a great one. Thank you so much, Sowmya, for joining us. Everyone certainly reach out to Sowmya and Lumera and learn more there. And thank you so much for joining. Thanks for having me. Of course. Okay. Alright. Well, before you guys all head out, once again, so much, thank you so much for joining today. Also, since Sowmya wasn't able to get to your questions live today, we do have a live AMA ask me anything session with Sowmya on May 29. We sent the link in the chat if you wanna register there. Additionally, we have a special gift. So if you want a one on one consultation to see how Rippling spend can be personalized to your company, you can also click the first link in the chat there or scan the QR code. In exchange for your time for signing up and attending a Rippling Spend demo, you'll get a free YETI cooler for the incoming summer months. So, again, this is available for both those who are completely new to Rippling and existing Rippling customers who are not yet on Rippling spend, but maybe you have something like Rippling HR or IT. And once again, in case you missed in the beginning, Rippling spends a unified platform to manage all your company spend including expense reimbursements, corporate cards, payroll, and bill pay. And its powerful automations and consolidated nature, has been able to help companies like Andros save a hundred thousand dollars per year on staff hours and software licenses, dropping their close time to just three days. And Pepsi of Worcester, who closed their book seven times faster now with a % compliance, and Victoria Beckham Beauty, who doubled their monthly spend, and they're able to still analyze and track everything in just fifteen minutes. So once again, here's the QR code, link in the chat at the very top. And again, we'll send everyone who attends the demo a free Yeti cooler for your time. Thank you all so much for coming out today. We really, really appreciate it. I hope you learned something new. And again, if you have more questions for Sowmya, we saw a lot of questions about things like privacy, use cases, and things like that. So definitely check out the AMA, and we will see you all in the next Rippling Spend webinar. Thank you, everybody.