Name: Beyond the Buzzwords: Tactical AI Use Cases in Finance
Uploaded: 2026-01-06T20:46:06.810Z
Duration: 1 h 12 s
Description: Beyond the Buzzwords: Tactical AI Use Cases in Finance

Transcript for "Beyond the Buzzwords: Tactical AI Use Cases in Finance": Thank you so much for joining us today. There's a lot of excitement about AI, but so many folks are saying like, what are some real use cases that finance and accounting teams are using today? So I'm thrilled to kick off today's discussion with two of the top experts in AI and modern finance. Quickly on me as the moderator, John Glasgow, founder and CEO of Campfire. We're the modern accounting platform for midsize companies. Soumya would love a quick background. Thanks for having me, John. Hi, everyone. I'm Soumya. I recently founded a company called Lumira to help finance teams deploy AI. Prior to this, I was controller at OpenAI. I was super lucky and got to join the company right as Tejucati launched and saw a really hyper growth, intense time at the company, but also learned a lot about how to use AI to supercharge our own work. Really excited to share a lot of those learnings here with you all today. Dave? Yeah, I'll go next. Hey, everybody. I'm David Wiesenack. I am a Finance and Accounting Professional. I work at Ramp as our expert in residence. We use our own platform, Ramp, internally, so I get a chance to deploy that and do a lot of product advisory with our team. Great. And before we hop in, Somya, quick exercise, we'd love for you to go through that. And while you're doing that, we've got a poll. We would love for everybody to fill out that poll. Today's session, you're going to see a lot of real world examples that, between the three of us, we've either used ourselves or we've seen our customers use. It tends to be a lot. In my opinion, one of the most effective ways to take the most out of this session is for you to start a notes doc now or have a notepad beside you. Write down examples you see that really resonate with you. If something comes to your mind about, oh, I do something really similar at work, I would love to try this out for myself, jot those ideas down while they're fresh in your mind as they pop up. And then after this call at the end of the call, we'll come back to this. And I think it will be a really good way to kickstart your own journey in AI adoption. I love that. It's such a good idea. I actually have my AI notetaker taking notes for me for the call right now, so I don't have to write it down. Love it. Love it. Alright. Well, let's take a quick look at the poll results. And so we've got 6% not using AI yet. 29% just starting. 50% looks like about half are testing the waters and we got some power users, 31 of your power users. So maybe we'll have you on the next panel that we host. But with that, I think irrespective of where you are in your journey, we're going to have some great takeaways, whether you're not using it all or you're in that power user category. So I think the first one is just to be super tactical, where are you both seeing AI be really good? And Somya, I'll start with you. I'd say right off the bat, a lot of accounting tends to be structured, especially once you're in the transactional accounting world. And when you're looking at structured data and you're using AI to write SQL or Python, I'd say the code generation capabilities are really, really good for finance use cases already. It is also a great learning tool. If you're new at a job and you're trying to get up to speed on all the things, you can really accelerate your own learning using AI. I learn a lot better in a Q and A format compared to reading a textbook, for example. So I think you have to do what works for you, but those are really good. I'd say anomaly detection and data analysis capabilities are really good today. And of course, the thing that we were all really impressed when CHADGBT first came out, it's really good at writing things. And I have to say, I definitely are on the side of lazy when I get started on things. And I'm much more motivated to work on something that's already there than start from a blank page. So anytime I'm drafting an ML, a policy, any doc, I usually ask ChatGPT for a first draft that I can then go edit and clean up. I find that it's a lot faster to edit than to write from scratch. So that's a lot of the ways that I've seen it used both by myself and by my team, but would love to hear your guys' perspective. Go for it, Dave. Yeah, the places where I'm seeing AI need to get better is full autonomy. So being able to give an agent a task to go off into the world, do work for you, come back. It's still early days when it comes to full autonomy, but it's very exciting to see where the future is going. And then, of course, the full business context. If you just go into Chechiuti or Claude or Perplexity and you ask it a question, it may not know the full depth of your business, who you are, so you of course have to prompt it, but it may not have all of the knowledge that gets passed down through from senior to junior on the team, all of the historical context of your company that you may be holding onto your head. So unless you give that full information, that full business context to your AI, it's not going to know, it's not going be able to react on that the way that a human can today. But I think that's going to get better over time as we all get better at prompting and giving it that full business context. Yeah, makes a ton of sense. To Samia's point, I found that because there's so much code on the Internet that the LM is largely trained on publicly available data. It's probably why it's so good at SQL. There's so much SQL online, but no one's posting all of their financial data on the Internet. And so there has been more limited data for it to really learn how to number crunch. And then to Dave's point, some of the multi step reasoning, we've seen some struggle and it'll miss a step or it'll just stop in the middle of a five part process. And so still some work to be done and some of the deep research models are helping with that, but still early days. All right. Well, Dave, I'm gonna start with you on this one. And I know this group is really aligned that there's two views here. There's the conversational interfaces call like a chat GPT that seems to have turned into a noun. And then we've got the product integrated AI where you're logging into existing tooling, whether it's Notion or Ramp, you guys have done a great job with AI and really using AI where you're already going. Would love for you, Dave, to talk through how you think about each one. Yeah, for sure. Usually when I have an ad hoc task or, it's something that maybe I'm doing for the first time, I'm going to start in that ChatGPT kind of, experience where there's a box with it and there's multiple models, but we'll use ChatGPT as an example. It's a box and you can just write anything. So I'm going to prompt it. I'm going to tell it a little bit about who I am, who the ChatGPT is, give it a persona, give it a task, let them understand what I'm trying to do, and then of course, what the output that I want from it. And that's going to be great for exploring a new idea, doing an ad hoc task. Maybe I'm uploading data to give it more context or something that's going to do. But when I think about more scalable features, that's where I think about AI that's built into a product like Ramp, Campfire AI, some other tools that you have, because that context is embedded, right? The job to be done is embedded into the product. And so it's really software that is supercharged by an AI workflow within it. And that's gonna be much more scalable. It's gonna be using the data that you have in the application to get that job done. And of course there's something in the middle, which is custom GPTs or the projects that are in perplexity. And that's where you've pre coded or pre written down a GPT on or pre prompted it rather. And you can use that one over and over again for a specific task. And you can share that amongst your team or your friends or your group. So everybody can utilize it over and over again, same thing. And that's somewhere in between in the middle here. Yeah, I'm seeing a little bit of convergence. And we'll see it in campfires that we have our own conversational interface. But then to your point, there's the JACCPTs of the world are moving into contextually aware and business contacts. So something in the middle is forming from both sides. Somya, would love your thoughts. I actually resonate a lot with what Dave just described. I think the I'll jump into a demo a little bit down the road to show you what I mean. But I think the barrier to entry for ChatGPT is really low. You don't need to have there's no learning how to use the tool. You can just ask it what you want help with. And I think the cool thing about that is if you have an idea or you have a problem, you have a really gnarly Excel file, you can just start talking to it about what your current process is, ideas for how to make it better. You can ask it for Excel formula help. You can say, what if I don't want to do this in Excel at all? Can you think of other ways I can do it? And I think one of the best ways to use these general purpose tools is to come up with a better mental model of how things can be improved. In my opinion, it's really hard to beat these models for that rapid prototyping and feedback. In my early days at OpenAI, I tried to have ChatGPT be the flux analysis writer for our P and L. And it can do a decent job. I think where that transformation happened for our team to we still use both, right? We use ChatGPT on its own, but we also bought software tools that have AI capabilities natively in them. And I think there's a few different ways to think about it. I'd say one is if it's a high transaction volume and you want all the other capabilities, you want the collaboration, you want it to be integrated in your process. You don't want to be downloading PDFs and uploading them back somewhere else. I think when you think about what that end to end process looks like, there are going to be places where it makes sense to do ad hoc versus there are places where it makes so much more sense to be embedded into your core. And that's another way to think about when you make that switch from ad hoc to a repeatable process. Yep, makes a ton of sense. Well, I know you mentioned you're going to be demoing, so you are up first on this. I'm super excited. I have not seen this yet. But Somya is going to take us through a couple of live examples of how they were using it at OpenAI within the accounting team and also how she advises folks following along today as well. So, I'm going to stop sharing and over to you. I will take it over. I am going to show you guys. Do you guys see my screen? Yes. Okay. The first example I have for you is imagine you just joined a team and you are tasked with six zero six. And you are one person and go to market has 200 people, then ratio is definitely off. You just need someone to help you with when order forms come, you want help speeding up your analysis. So this is you can tell this is actually a ChadGBT generated order form. So how's that for a meta demo? But this is an example order form. You will of course do this with your own company's data. And then I want to use the audit firm's six zero six guide. And this is imagine you have like a special document either about your industry or something that recently came out and you want to make sure that it's incorporated into the analysis. So the way to get started using ChildGPT is to simply say, I always like to start my prompt with a persona. Don't ask me why this works. It just works. If I tell it you're a CPA with expertise in this, it just tends to do much better. Try it if you haven't already. And then all I'm doing is I'm going to say, I will give you an order form and my audit firm's guide. Review this. Identify nuances. This is another thing I really like putting into my prompts. If you see ambiguity or you see things that are not defined and you're not sure about something, please don't guess and just make up stuff. Just tell me or put it in a list of open questions at the end. This is a way that you can cover your bases. A lot of people say, What do I do when the model hallucinates? This is kind of a guardrail against that. I have to say hallucination is getting better and better, like it's happening less and less. But I would just much rather be sure than not sure. So this is kind of how I structured the prompt. You'll also notice that I'm kind of doing this prompt before I'm giving it to docs. And I think it's a way for me to make sure that it really understood what I'm trying to do. And if you just upload the two PDFs upfront, it'll give you something. But I'm never sure if it really understood what I want. So the nice thing about doing it this way is now it knows I'm supposed to upload two things, the order form and the audit guide, right? And then it's saying, these are the things I'm going to do. And if it's going to say it's pretty detailed. One thing that is interesting is TragicPutin knows I like seeing things in table form. So it gives me this in table form. You can put your custom instructions on the whole. I don't like seeing long lists. I just can't process it very well. So you can see that it's tabulating its steps for me. If you want it to do different things, if you want to say, I don't really care about step three, do these other things, you can give it feedback at this step. So then what I did was basically go back to this order form and the six zero six guide, and I just uploaded those two into ChatGPT. And it took a minute, which is not instant gratification, but it's about as fast as like, you know, this is not bad. It went through the different SaaS access implementation, tech support. It's looking at like options for additional users at prices. I would say this is pretty good. It's very fulsome. And if you're someone that just, this is obviously like a made up example, so it's not as complicated. But if your company has really complex revenue arrangements and you want something to break down by the type of thing you're selling, by each performance obligation, and do every part of the analysis for you, this is a really good starting point. Now, you can keep tweaking this until you get it to the output you want. Maybe you have a template for how you document the conclusions for every single order form. You can give it that template and say, I just need you to fill out these columns for me. This is what I mean about the open questions and information gaps. This is really helpful because you now know, one, for yourself, are there questions I should be going back to the sales team or someone else with? Is there feedback I need to give legal because this is not clearly worded and this could be a problem down the road? But this is an example of how you can accelerate your own work for something as as routine in your job as documenting conclusions for this. You can also use this for non like if you have, say, a lease and your company doesn't enter into lease agreements all the time and you want to do a deep dive, but you don't think of yourself as an eight forty two expert, that's Okay. You can follow the same process and really just upskill yourself in a domain that you're not super comfortable with. So you can do this one off. You can upload this, upload every order form that comes your way. The way to go the next level sorry, I'm trying to figure out how to move this Zoom window thing. Okay. What you're not someone like, what I love about your prompt is it's pretty much what I would like, the instructions that I would give someone junior on the team to tell them what I want them to do. And if you don't know about what prompt to do, you could also just ask ChatGPT, like, help me write a prompt. It'll ask you questions. It'll then give you a prompt and you can copy and paste that and get started now with that new prompt. Yeah, totally. Dave, you mentioned GPTs earlier as kind of an in between of how do you go from the general purpose box to a templatized version of this. So for those in the audience who have not used GPTs before, I'll quickly show you how to click through. So you're going to go to ChatGPT, go to Explore GPTs on the left. And then there's a button here to Create GPT, and it takes you to this flow. For this demo, I actually have a GPT that I created called six zero six Order Form Review. I'm going to walk you through the prompt and how it's set up. A lot of people ask me if they need a technical resource to do this. And I'm here to show you, you actually don't at all. You can create a GPT the same way you just talk to ChatGPT. I pasted a prompt that's really similar to what I had my own instance of chat, right? I said you're a CPA, six zero six experience, analyze contract terms, ambiguity, don't make up stuff, like all the same stuff. Except you only need to do this instruction once. So you're setting this prompt and these instructions and you're saving it in a way that's reusable, both for yourself, but also for anyone else in your team. And then I've uploaded my guide here as knowledge if you have other things. So earlier I mentioned, if you have a template you wanted to follow, you can upload that here. If you want to be even more kind of guardrailed around what people can upload, you can upload a sample order form. And if a user says, how do I use this bot? It'll basically say, this is an example of an order form you can give me. So you can go as intense on this GPT configuration as possible. But only one person in the team needs to really do that. And everyone else that's doing it, all you're going to do is go to this GPT in a new chat. They're going to drop this order form. And they don't have to write the prompt again. I'm not going to say anything else. Here goes an order form. It's searching its knowledge. The knowledge is the guide that you uploaded for it. And then it's doing this for you. So this is an example. I use 606 because we were just talking about it. But we've used this type of thing for deal desk guidelines for the sales team to know when they can discount, when they can make exceptions, who they need to talk to about exceptions. You can put all of those in a GPT so that they can just kind of interact. I've made a GPT about our team T and E policy. So people constantly want to know how much can I spend on a team social? And like, those questions are handled really well by GPTs. AP coding, if you want to put your AP coding business logic in here and have new people on board using a GPT, you can do It's really flexible and customizable. I'm going to switch gears a little bit and show you something that is called deep research. It's a little bit of like a confusing term for accounting and finance. You don't think it's something we can use. But let me show you an example of, say you want a footnote disclosure for debt covenants. And you can ask RGPT without doing deep research, and it'll give you something. Here's a deep dive. Okay, great. Right? But then you don't know where exactly it came up with these from. It's hard to trace back. You kind of have to trust it or kind of, you know, pierce it by yourself. Deep research is literally just a button that you click when you let me hang on. So in a new conversation, you're going to ask the same prompt here, and then you're just going to hit run deep research. So that's how you get started on it. So what I did here is I gave it the same exact prompt I put in my other chat, but then I used the click deep research button. So the first thing you're going to see is it doesn't just give you an answer. It comes up with clarifying questions. This is a good sign. These are things that the tool needs to know about what I'm asking. I usually copy paste the questions and just respond in line like I would to a human. And then this is the other feature of Deep Research. This thing took twelve minutes to complete, which is another reason I did this before this webinar and not real time and make you all wait for twelve minutes while this happened. But it did 127 searches and found 22 sources. This is one of my favorite things about Deep Research because you can now trace back to all the different things it looked at. I gave it Microsoft as a ticker so you can see that it's pulling things from their own IR website. And the quality of the disclosure is just much better in my opinion. It's also linked and sourced, so you can kind of trace back. So if you previously used a tool like JWT and you were kind of like, it's not great for technical accounting and research, I really encourage you to go back, try the same thing with deep research. It just takes it to the next level. But I know I'm running up on time, so I want to stop sharing right now. But that's a little bit of a primer to how you can start using this right now. So you're getting a lot of really great questions here on chat, we'll have to answer some of those. Thank you for not making us sit through twelve minutes while I did the research. Definitely. One quick one, Somya, is a lot of folks are asking which model were you running there for the first example before you got into deep research? And then second was, when you are running deep research, was that a higher tier of the paid account? Okay, great questions. One, my favorite model is the old series of models in Charge GPT, which are the thinking models. So the one series, right now I use three, and it's my personal favorite. It's supposed to be for really analytical thinking and coding and data analysis. But I find that it's actually also really good for accounting, especially grabbing the nuances that are not spelled out in the open. So that's the model. Sorry, what was the second question? Around you were using deep research. Was that like on a higher tier? I am on a pro account, but I believe it's available on other tiers as well. So don't you know, it's not like kind of feature gated. You might get fewer queries a month, but it's available on other tiers as well. And just to clarify, we here at RAMP, we have an enterprise agreement, so we have data protections. So before you load in, if you're on the free version of ChatGPT, or even the pro version if you personally buy it, you're not going to you may not want to load in confidential information, your company's financials, things like that, because your company has not negotiated the rights of what you're loading into ChatGPT. So if you do load in financials or policies or anything that's confidential, make sure you're on the enterprise version and your company has kind of negotiated with OpenAI or the other programs. They all have the similar type of things. Great call, Dave. Deep research should be available for the free version. You just might only have, let's say, five queries per month or something like that. Yeah. They give you just enough to know what you're missing out. Yes. Awesome. Well, let's hop over and we're going to transition a bit. Dave's going to kick off the product integrated AI examples. Yeah, of course. When I'm using ChatGPT, like I said, the box may be overwhelming, right? You have to start, you have to add in persona, you have to prompt it. That may be a little overwhelming and of course it's not integrated with your workflows. So whether you're on Ramp or you're in your ERP, it doesn't have all the context of what's going in your financials. So here's an example of RAMP where we have AI that's integrated into our product. Here you can see someone doing their expenses. And this is just an SMS, but this is similar to on email, in the web application or in the mobile app, where when you start doing your expenses, we are going to do things like suggest a memo based on what we can tell from the receipt or from your calendar. If you have Google Calendar integrated. We are going to auto generate receipts based on the card details, if you have that feature turned on. We're obviously going to do OCR. OCR has been around for a long time, but now with LLMs, we can do a much better job of understanding that receipt. So it's really supercharged because of AI and LLMs. And then of course, flagging transactions that may be inappropriate for a business purpose. All of that is AI LLMs that are integrated into the product that you don't need to prompt, you don't even need to think about. It's kind of just software that gets supercharged. Love it. By the way, super impressed I was able to read that crumpled up Chipotle receipt. So that's off to your team on that. That's the AI doing its job. All right. Same thing on on Bill Pay with with invoices. I actually wanna showcase something. This came out from AppSend, which is a competitor, but it came out, this week. They have this new feature called AppSend inbox. And so this is going to be integrated to their platform, their bill pay accounts payable platform, and able to read all of the emails that you as an AP team will be receiving. And of course, drafting responses, right? Being proactive and drafting a response back to a vendor that's asking for an update on when they're gonna get paid, or maybe processing that invoice or processing that W-nine for you. So really cool features coming out from AppsZen, and of course, more of that will be coming out from Ramp as well. And then accounting coding, right? We talked about if you have thousands of transactions that are coming through for your company, someone on your team, a bookkeeper, a controller, an accounting associate are going to be coding those transactions, right? 99% of all of your transactions look the same, right? It's the same software vendors you're working with, it's the same expense reports, people getting coffee and dinner with clients, things like that. Most of your transactions you've seen before and you've already coded up. The future, and while we have this here on ramp, the future of a lot of bookkeeping tasks, is that because we've seen those transactions before, we can anticipate how you as an accountant want to code those. And anything that the AI is unsure of, it's gonna ask you, hey, how would you like to code this? And then remember that for next time. So being able to use all of the transactions that you have in your ERP, how you've coded them in the past, any memos that you have in terms of how you want to code things, just as you would teach a junior accountant on your team, here's how we code things, you are also going to be able to teach AI and it's going to learn and be able to do that for you in the future. So we have a lot of that in ramp and there should be a lot more that's coming, it'll get better over time. And then lastly, we all use Excel, we all use Google Sheets. If you are not experimenting with the new AI function in Google Sheets, you should get started. It's still early days. There's some things that it can and cannot do. But now natively within Google Sheets, a tool that we are already using, you can bring in AI. All you have to do is just equals AI as a function, and you can start using AI within each cell of Google Sheets. I'm sure Excel will have this as well. And so this is a way where you can use AI embedded into a product you're already using, not needing to kind of copy and paste it into ChatGPT, prompt it and then copy and paste it back, but rather doing it all Google Sheets and Excel. So really cool stuff. It's not great with numbers quite yet. It's still fantastic with words and reasoning, not quite numbers. So if you do need to do formulas, functions in Excel and you're not sure which one to use, go to ChatGPT, tell it what you're trying to do, maybe give it some sample data and say what are the best Excel formulas I can do to achieve this? Gets really good, especially if you really complex formulas or functions that are kind of like nested and maybe something's not working and you're like, you know, how do I debug this? Where is it going wrong? Throw it in ChatGPT, give it a prompt. It's the best Excel formula builder I've ever seen in my life. I use it all the time. But I'm excited for that to be within Excel and Google Sheets in the future. But for right now, I use a lot of that in the box, the AI box. This is a game changer. My first job in finance coming out of school, they actually took the mouse away from us and excelled it to be entirely keyboard shortcuts. Feels like this is really just changing the game in terms of VLOOKUPs and all the nested statements that we all used to do. Just and In it's going to go from the future, they're going to take away the keyboard too. All you're going to do is talk to it and vibe Excel. You're just going have to talk to it and do all the coding for it. Awesome. Well, one last slide here, and then I'm going to hop into a live Campfire AI demo. We've got embedded AI. So this is similar to what Dave described, like for what they're doing at Ramp, Cache Racks, and just throughout the product, it's performing a lot of AI activities, uploading a customer contract. It's just going to auto pull it in for you. But then we've got Ember AI, and that is our version of the conversational interface. And so it's flex commentary on a budget or on the monthly close. But we're gonna hop into the Ember example here and look at some real use cases with the live demo. I love that you called it Ember AI. That's such a good name for it. Yeah, we want to do something on theme. So within the Campfire family. What we've done with it, and it's similar to really any chat interface, is you just ask it a question similar to what Samyet was doing. I think one key difference is it's really auditable because it's sitting within your accounting software. And we can actually put these tasks into your close management checklist or really in other places of Campfire. And these unique URLs are embedded directly in your ledger. So it's always going to show you how it completes the task. And then it's always going to pull the actual source data. And because Campfire has all of your customer contracts, all of your GL sub ledgers, we never have token limits because we've built this in a way that actually it's going to infinitely run. If you want to do year over year flux, and you got 10,000,000 entries per year, it's actually capable of going through every transaction and running that level of analysis. And so when I was in finance, I was getting pounded with all sorts of chart requests. Now some of our customers are even giving their CEO and other folks this interface here, and they can answer their own questions. Or you can just go ahead and I know you wanted that chart and you only gave me twenty minutes to build it. Well, here you go. I'm actually able to do that now. Let's go through a few more examples together. So here's one. Accruals is not fun for those on the accounting side. Gunderson Detmer is our law firm. I said, just grab a three month trail. I can't wait for the bill to come in. Legal has not responded to my Slack to give me an estimate. So what it's doing is it's going and grabbing our three months of prior accruals for this law firm. And then it's running the average. And then what it did here is it drafted my entry. And I'll just click create entry. I can review it. And then I can just go ahead and post. Let's look at a more complicated example. Meeting with a lot of finance and accounting folks that are like, I get these gnarly AWS Marketplace bills. There's like 30 vendors on it. And what I did here was I took the bill and we an estimated statement, but the bill isn't hit yet. So I uploaded the PDF of the bill and what it did was it actually went and I'll just show you here. It grabbed the vendors on each row, Snowflake, Datadog, and I even sent it over to engineering here on this request. So you can see I'm like, send this to Eng. This is their spend. So let's make sure we code it right. And it drafted everything for me and read the whole PDF. And again, I can click create bill here and I'm going to review just like it's as if I created this entry myself. We can go ahead and post. Now that's for closing faster, but we want to close faster with confidence. And so a lot of our customers are creating a close task, just saying like audit my close. And it's going through all the transactions, all of your contracts. And we can see here what it found. Yes, there's a few uncategorized items, but I want to actually run a quick filter and let's just look for duplicate entries. And it's like, all right, let's take a look. I can just download this, drop it in Slack, send it over to somebody. And it actually went and found some things we want to clean up. Because we all know we're running super lean as an accounting team. And sometimes there's no one there that's really checking our work, But we want to be confident going into the next audit, that next board meeting, that next fundraise. And so this is how we're doing that for our customers. A more complicated example and in line with what Soumya did, is you can upload your six zero six memo into Campfire AI. And I uploaded ours here. And I said, go through all of our contracts. So, Soumya, I'm gonna spar with you a little here. Instead of looking at one contract, Campfire pulled all of our contracts. We've got hundreds of contracts in this account. And then it actually does, I had to do a sample. It went and actually found contracts that are non compliant. And again, I can go ahead and it looks like we've got some things to clean up. I'll go ahead and pull this out and we can meet as a team. And as always, it's going to be showing its whole work and all the sub ledgers are sitting here. And so when I meet with like big four auditors and they're like, yeah, you guys are really building this in a way that's getting us comfortable. Can I add one thing, John? This is actually a really good demo of when to do like kind of the box version or the ChatGPT version of AI versus an integrated version of AI. Because what I demoed works for a single contract at a time. But in reality, you have hundreds of contracts a month. And you want something that comes with the audit trails, the collaboration, the ability to data pipeline into the system. And I think this is just a beautiful experience for the team to not only get the job done, but also have all the other guardrails around it. Great point. Great point. When I was at Adobe Finance, we spent weeks taking marketing data and doing LTV to CAC and merging it with all of our data out of Salesforce CRM. So what I did here was I said, I uploaded our marketing dataset, the lead source for all of our customers, those hundreds of customers we looked at earlier. And then what I did was, I said, go ahead and merge that with all the financial data in Campfire AI. And again, had to just run a sample for us on most recent customers. And we can see here it grabbed customers, it grabbed everybody on the marketing dataset for their source, looked at total revenue, it looked at churn. And so we've got some decent insights. Actually have like high churn coming off of social media, something to dig into more. But look, I can just make up commentary or I can be like, Ember, what would you do? And so Ember is our CFO, and it's actually going and saying, let's go ahead and make some tactical decisions. Here's going to be the financial impact if we make these moves. Now I can go meet with our CMO, like when I was at Adobe, and I've got much more confidence in the recommendations that I'm making and really help them and really up level the role and make it more strategic. Because we're closing faster on the accounting side, we now have time to do more strategic work. And just going to quickly go through a few other ones. Head of Eng is like, Hey, how am I pacing? It feels like I'm a little behind on budget. Go ahead and ask it. It's just going to go ahead and tell you exactly how that's rolling along. And then another one is like, and one that's super helpful for me is I'll just take the last board minutes and I'll drop that in here. And then I'm just saying like, hey, based on all of our financial data, all the sub ledgers, GL, go through everything and going into the next board meeting, where should I be focused? And it's coming up with recommendations. Last board meeting, there was some concerns around our runway and burn. And so it's actually going in there and coming up with ways, like some talking points for me. I know we all prepare a lot of back pocket analysis. We get worried even in a board meeting, you can start querying your data in here as you get into that rabbit hole with a board member and you don't want your gear in headlights, we can help you get through those moments. Sean, where did you load up your board minutes? Because your prompt was up at the top. I didn't see it had any do you need get up? Yeah. I'll fire it up from scratch here. So what we do is you can go ahead and you just drop in your your board minutes here. And then so we have this upload button. And then it'll just go ahead and run here. John, I've I've seen an early version of this a few months ago, but this is wild. Holy moly. This is amazing. And then all your chats are saved, unique URLs. But, so all the work we're performing, I can go grab something and you can see here's all the work that was done. So but that's a quick campfire tour on our AI. Really cool. Next up, AI maturity model. We heard at the beginning, there's some folks that are like, hey, doing none, and we got some power users attending today. Soumya, would love you you did a great job taking your team at OpenAI from Explore to Transform. Just would love to hear your thoughts on this. Yeah, I think my approach has always been find the biggest process bottlenecks that exist in our work today, and then try to find solutions for those. Because I think people sometimes do this backwards. They're really excited about AI and trying to start from the use cases and then trying to wedge that into their work. But I think if you truly want to make that transformation, it's much more rewarding to start with the most painful things. So my first use case was actually building a Python script using AI to automate our GPU accounting at OpenAI. You realize that there's no off the shelf SaaS tool that's going to really do GPU accounting at the scale and complexity that we were doing. That company would have one customer, us, and this company doesn't exist. So we really had to take it from this ad hoc exploration to deploying a Python script that would automate our own dashboarding for that problem. The nice thing about picking the thorniest problem and then trying to put AI into that and making it better is you can show the team really impactful results. So if you can shave off your closed timeline by five days, that means you've freed up five days of people's time to go do other projects in transformation. And I think instead of going for the small incremental wins, go for the big meaty projects. And as crazy and hard as it is, you'll find that with AI, you can actually get a lot more done than you previously thought you could. My whole thing is shorten your close timeline, create more time that's not closed so you can work on these improvement projects, transformation projects, system implementations. And then that puts you in this very positive flywheel of you have more time to save yourself your future self more time. We did start with ChatGPT for most things, but over time, like I mentioned, for Flex Analysis, we put in a tool that synced with NetSuite, which was our ERP, to kind of do this at scale. Similarly, for order form review, so the tool we implemented was numeric. We implemented Clarity for revenue process and deal desk and order form review. And I think once you do this prototyping, you get a really good sense of where do we need to put in an AI enabled tool that will give us the guardrails and the integrations that we care about. I love it. And one thing, by the way, that I think really stood out there was each of your tasks as a team member, you might be in different phases. So when we talked the other day, you made a great point that you are always pushing the team between closes, experiment on what tasks took the longest, and then you explore. So maybe you're in transform for a few tasks. You've got it fully automated, but then some of them are still manually doing all this in a spreadsheet because we figured it out. And so then you've got these windows between closes to go, how do we take the task that's currently taking the longest and put it through this framework? And then that's gonna really allow you to then move it transform. To And then you move on to the next task for Explore. Dave, anything you want to add? No, I agree. I think personally, I'm more in the experiment innovate place. I think as a finance team here at Ramp, we are very much in the innovate section for a lot of functionality. I'll say some of the younger people at Campfire AI, I watch them work with AI and certainly ahead of where I am. So learning from your peers, I think at the bottom here, got some tactical advice around looking to the folks in the team that are really pushing the boundaries and even just watching them work. I'll go over to our engineering team and just watch them use AI. I'm like, wow, things are really moving fast. Somya, again, want to make this super tactical. So we'd love for you to just talk through this framework for the group. Yeah, Dave alluded to this earlier, but I want to reiterate, for this to be really effective for you in using AI at work, make sure you're in a place where you're comfortable, your IT team is comfortable with enterprise security and data privacy. My recommendation is to get a work account with your work email and opt out of training on your data. This is not super publicly advertised, but you can actually opt out of data training even on a free account in GPT. You can opt out of memory. This is not like a P walled feature. You just need to make sure you go into your settings and train. The other thing that people think the models are doing with data is training on your actual prompts and answers. It's actually just using your thumbs up or thumbs down to make the models better. It's not actually going to put your data set into the larger training data. So I'd say take the time to get educated on what's really happening with models, but it's worth it for you to get yourself a Teams account or enterprise account. The barrier to getting started is a lot lower than you probably think today. So that's one takeaway. The second thing is remember that notes file we asked you to get started and start jotting down ideas? Go through that at the end of this call, take five minutes and circle one that you're going to go try out for yourself. Maybe it's not going be perfect the first go around, but you'll learn a lot in that initial iteration and just exploring phase. The best way to use this is to actually just get started. It's infinitely patient. It doesn't get mad when you ask too many questions. So don't hold back. Just go for it. And the last thing I'll say, this worked really well. At OpenAI, we had people on the finance team who were very excited about AI as a technology, but we still had a lot of difficulty in getting people to really use it to accelerate their own work. So I found that creating these exploring places for people where it's low pressure, you're not tying it to performance, and you're just saying, let's come together in a team offset. We'll carve out an hour for hackathon, get together in groups of two or three people, put some ideas forward. And your idea can literally just be a GPT. You don't need to build some very complex agent orchestration or any fancy words. Start by prompting. And then if your prompt's working really well, turn that into a GPT. And that's a really fun thing to share within the team. We continue to hold these quarterly team off sites, sometimes on-site, off sites. It's very low production, but it's very, very effective in getting people more and more comfortable using them. Amazing. Thank you for sharing that. We're going to do spicy takes. I want to leave time for Q and A. So we're actually going to do one minute of spicy takes. So I'm going to start the clock here. And when it hits four, we're going to move into Q and A. So folks, I know we got a ton of questions. Now's your time to get them in. If you have one. AI OKRs, quick hitters here. Dave, I know you have a strong opinion. Go. Yeah, I think it should be part of your evaluation. So you should have OKRs, your team should have OKRs. And as you evaluate your team, make sure they're using AI at least for 1020% of their evaluation criteria. And for folks, objective key result is OKR. It just means goals for people on your team. Somya, thoughts? The reason I think this is a spicy take is when you put it on an OKR, you're tying it to individual and team performance goals, right? And this directly impacts how you're getting rated at work. My view on this is when you're in the early stages, when you're exploring, this puts a lot of pressure on the team. And the thing with AI is it takes a lot longer than you think to actually get it to fully work. And when you put an OKR deadline on it, I think it has this effect of people just, they sandbag their goals. They don't shoot for the moon. They just say, well, I'll just try this one little thing, which is definitely going to work, but it's not going to have the biggest impact. So my hot take is when you're in early stages, give people the freedom, make it not pressure and fun and see kind of the ideas generating. And when you're in that transform and you're kind of climbing up that maturity model, then make it more kind of wholesome. But it kind of depends on the team culture and the company culture too. Okay, awesome. Well, we're going to skip the rest. Let's move into Q and A. And let me just move this along. Okay, so we got information for us if you want to ping anybody directly. First question here, how can you determine which AI is the best fit for your needs between Claw, Gemini, Deep Sea, Perplexity? Anyone have a quick take? So, you're probably a little biased on this, but if you have any candid thoughts, either of you? Yeah, I would say try multiple models out. They're all slightly different for different things. They're all getting better every single few months, every single week. So try them all out or at least the ones that you have access to. Right? If you're a Microsoft shop, you probably have access to Copilot pretty easily. If you're a Google shop and and you use Gmail, you probably have access to Gemini pretty easily. If you have an enterprise account for ChatGPT or an enterprise account for Perplexity, use the ones you have access to, but try all of them, whether that's for work, for at home, and experiment with different types of models, not just language models. Experiment with video creation or audio creation. There's a lot of really cool stuff out there. Just play around, test it, experiment, and try them all. Awesome. And yeah, to Dave's point, almost all of you are going have some sort of enterprise license. So, folks are forced into a Gemini or Claude or ChatGPT. Someone said, I previously tried to create a conversation with ChatGPT, fed it in our info, can populate customer onboarding forms. It wasn't good at editing PDFs on my behalf. What are their time saving tasks to use a controller to use QPT-four? I know we covered some of those, but just general guidance on like, Somi, do you keep iterating when it gets stuck? Or do you just like, when do you tap out? Like, hey, maybe just AI is not ready for a task. I think, in my opinion, when you give it a full end to end task, it doesn't do a great job. So when I get to that point of I'm hitting a wall, my first instinct is to say, Okay, how can I chunk this into smaller tasks? And then I'll try to automate the thing that is the worst part to do manually. So for example, if you have like 100 PDFs and the thing that's slowest is copy pasting each value, for example, maybe the fastest thing is to have it write a Python script or some other script to take those PDF data and drop it into an Excel file. Because once it's in Excel, I can do some of the other stuff manually pretty fast. So it's a little bit of like, don't try to drop in a 10 tab, really complex multi formula Excel file and expect it to get it right on the first go. Try to say, Okay, what is one tab that works? And let's build from the smallest thing that works and then build upwards. Great. A lot of questions about data security, even though, yes, your data is not being trained. And yes, there's been some hiccups where someone was being shared other person's data. Super quick answer, either of you. How are folks comfortable with data security for AI? I utilize the data security team here at Ramp. They've opted us into enterprise agreements, and I feel confident that they've done their review as a company. Tamiya? Definitely, I'd say work with your IT team, work with your security teams. Usually, the type of plan you're in makes a lot of difference. The other thing I do, I am very careful with sensitive data. Like payroll data, even though it's ChatGPT at OpenAI, I don't even want it to accidentally be visible to anyone. So I usually use dummy data. Imagine you're trying to do SQL or a Python script. You can give it 10 rows of fake data and just a column header names. Turns out column header names is all it needs to write your code for you. The dummy data is just for you to be able to validate if it's doing it right. So if you're really worried about data security, don't let that hold you back. If it can do it right on dummy or like fake data, you can then take that code and run it locally or run it in your own secure environment with real data. So I feel like a lot of the barriers have workarounds. And the creative thing is to be able to say, if you're worried about this, don't let that be the thing that stops you. Just think about, Okay, how can I get around that? Maybe thinking about it a slightly different way. Awesome. Quick one for me. Is Campfire AI like a full GL, a general ledger replacement? Or how does everything I demoed sit on top of it? Campfire AI is a full replacement for a NetSuite, an Intact, a QuickBooks. So it's our own NetSuite and our own AI engine. We found it just works much better when we are owning the end to end process. And so think of us as being a full replacement if you're on a system like NetSuite. Next question was, any AI resources podcasts that you recommend for learning? Am I allowed to plug Yes, of course, plug. I just started Lomera to really help finance teams get over this hump of deploying AI, both just start using TrackAPT, Gemini, FOD, but also go adopt AI enabled tools like RAM, Campfire AI, Clarity, Demerit. There's many that I have personal experience with. And my goal is for people to hear from each other in the industry on how things are going so we can adopt. So sign up for blog, Substack, and also happy to share some of these resources with all of you. So if you write a lot of good content, just at least check out her blog. Dave, go for it. Yeah, I just posted a link to an OpenAI guide that I've read and I really like and I find myself referencing a lot. OpenAI, Perplexity, Google, they all put out lots of materials on how to use their tools. They're very approachable. Usually, anything they're explaining takes five, ten minutes to try. And I learn a lot from those resources. ChadGBT is putting out a resource. I'm going to read it and learn more about it. I put the link just in here. If it's not working, Leah, I'll try to send it you. The other thing is I often find myself just asking ChatGPT or whatever system that I'm working with how to interact with it and how to get the most out of it. How can I prompt you to do X, Y, and Z? And it'll kind of tell me how to work with it. So it's another way. If you just don't know where to start, ask it. Love it. ChatGPT's new memory feature. We got two minutes left. ChatGPT's new memory feature. How are you using it? Do you find it helpful? I think also the custom GPTs that you showed earlier is probably a bit of a hybrid where you're giving it specific documents from a memory standpoint. But any thoughts on memory was one of the questions. I think memory is a little bit of a hurdle you have to cross for yourself personally. I use ChatGPT. I'm a multiple times per hour type user of ChatGPT. And for me, having memory is really helpful because then I don't have to keep telling it things like, I like things in table form. I don't like things in overly flowery wording. Just tell me the things quickly so I can move on. And I think it just understands my persona and the things I care about and the things I've asked for it before. So I think it just gives me better answers. At my house, my husband does not like memory at all. He prefers to have it be like a brand new experience every time. He clears all his chats on a regular basis, has memory turned off. So I think you can get value out of these tools with or without memory. It kind of is a level of how much trust are you going to put in the system. Well, know we're coming up on time. I just want to thank everybody for joining. This was by far our most attended webinar ever. Thanks for hopping on. We're going to send some follow-up resources. And again, contact info is here. Please ping me, Somia, Dave, any follow-up questions. And thanks for the time. Have a good one. Thanks, everybody. Thanks for having us. Thank you.