Name: How Meta Wrote the Rules for AI in Accounting
Uploaded: 2026-08-26T18:16:31.124Z
Duration: 58 min 48 s
Description: How Meta Wrote the Rules for AI in Accounting

Transcript for "How Meta Wrote the Rules for AI in Accounting": Hey, all. Thanks for joining us today. I was gonna give everybody one minute quickly to get settled. Good morning. If you're on the West Coast, good afternoon. If you're on the the East Coast and, nighttime and around the world. So thanks everybody for joining today. We'd love to hear where everybody's at, so feel free to drop it in the chat, where you're calling in from, and, be sure to get ready with questions. We'll have some live q and a with Aaron at the end. So if you have anything as we go, feel free to to save those questions and, we'll try and cover them live. Alright. We're gonna go ahead and get started. I'm John Glasgow, founder and CEO at Campfire. We're the AI and AWRP for modern teams. And, I had a ten year finance career before founding Campfire and was always frustrated with our ERP. And so I went and founded Campfire, but the big unlock for us was AI. There's a lot of opportunity with AI to take everything from invoice matching to accruals to financial reporting and rethink it from the ground up. And so that's what we've done with Campfire, and we're honored to have some of the fastest growing AI companies like Replit as customers over to air aerospace manufacturing companies, rocket ship companies, professional service firms, nonprofits, and everybody in between, like physician groups as well. And we have customers in every continent. So irrespective of where you're you're calling in today, we would love to chat. Aaron, would love a quick intro from you before we get started. Well, thanks, John, and and it's great to be with you. I think, I can speak. We're hearing a lot about what you guys are doing and what you're building and how you're rethinking how ERP is is done. So it's actually exciting to to be part of this and, and and really just follow your journey, as you try to rethink the way finance is done, which is probably a theme for this conversation. Yes. But yeah. No. Happy to happy to join. So my name is Aaron Anderson. I'm currently the chief accounting officer at Meta. Kind of started my my journey, a proud graduate of the University of Nebraska. Spent ten years kind of in a firm, both in kind of audit, national office, global office where I did a lot of IFRS. If we have IFRS people, great to hear from you. I still love my IFRS. And then spent some time at, IBM, Mastercard, and PayPal before, joining here about three years ago. Amazing. And Aaron and I met, quickly at a at an accounting dinner, here in the Bay Area. And since then, we've chatted a little bit and honored to have him with me today. Before we hop in to questions with Aaron and call it a virtual fireside chat, we're gonna do a quick poll. So we'd love to see where everybody is in their journey. You should see a poll pop up at this time, and we would love to hear from you. And there's no wrong answer here. So I think this helps in, influence the conversation today as well. So please chime in. I will give everybody another few seconds to answer the poll. And let's go ahead and take a look at the results. Let me see the results. K. Well, they're not coming up on my end. Give me one second. Oh, here we go. So we've got 51 votes. It's part of our day to day. And 35 votes, we're piloting and experimenting. And three folks, not yet, but planning to use AI. And so wherever you are on your journey, we're here to help you take the next step. Okay. So it sounds like most folks are using it part on the day to day. I think a little bit of this discussion will be how to how to, how to really go a lot further than you are. I think a lot of folks are saying I'm like, I'm using it and and it's been great, but what does, like, max velocity look like for AI? And, I think Aaron is certainly on the on the front end and and doing it at scale. So, Aaron, I I would love to start, you know, I I I meet with so many finance and accounting teams and they say, well, we're big for audit or publicly traded. You know? Like, that's cool. Like, AI is great for the the smaller accounting teams. You're at, an you've got a quite a large operation and I think operating in the public markets. And would love to hear, you know, for anyone that's skeptical on AI, using it at scale. I think people say, what if it hallucinates? You know, what what would you tell them for for folks that, wanna know more about using it at scale? Yeah. It's a great question, and I'm probably, I'm gonna give you somewhat of an odd answer, which is I think it's right to be skeptical. And I don't mean I I think it's there's obviously a fine line between skepticism and, you know, negativity. I think anytime we adopt something new, it's critical for us, you know, in our profession to say, okay. Our job is to present the truth. Our job is to to verify that truth. As I implement any technology or any capability in into my ecosystem, you have to have that healthy skepticism. And and I think it's one thing I I love about our profession is we just don't jump kind of at the next thing. We we we we test it. We understand it. We learn it. So I'd say number one, it's good to be skeptical. I'd say number two, though, is we have actually a pretty long history as a profession, and I think anyone who's been in this profession for, you know, more than ten years understands how to apply new technology. And we can talk about that later, but we we know how to do that. And so I think you you just need to go back to the skill set you have, which is, you know, how do I understand it? How do I verify it? How. do I make sure that, you know, if I'm signing my name to it, that I can be confident it's producing, you know, the right results? I I think my learning and we've been at this probably you know, in certain areas, we've been using machine learning for probably seven, eight years at this point, but we've had, you know, kinda your more advanced AI and more agentic capabilities kinda last probably two years. We've been year and a half, we've been really trying to do it at scale. And, you know, I think my learning is it it can be done. However, it takes a lot longer than you think. And, so don't believe any of those stories out there that say, oh, someone's gonna vibe code it over, like, an evening or weekend. and then, like, put something into production. It's just it doesn't happen. And I think that's been my learning is that to do it at scale in a reliable way, it takes, way more time than you think to do UAT testing than what we're used to, in when you're using AgenTic capabilities. And it takes a lot more, monitoring and maintenance and a new skill set to do that. So I do think it's great to be skeptical. I worry about the people that dive in and just say, let's apply a new you know, something new without the ability to really understand how to maintain it. So I healthy skepticism is the right approach. And, I completely agree with all of that. And you you've clearly been on this journey and you've thought about it a lot. I mean, I was talking to a reporter yesterday and they're like, hey. Engineering, AI's taken off and, like, hey. Sales has taken off. And they were like, when's the big moment for finance and accounting? And so I think there's this question of, like, it's clearly you're there today, but, like, more broadly, like, when and why, let's start with when did meta lean in on AI? Was it the was it the, like, 2023 kinda takeoff of of GPTs, or was it before that? Or Sharris went share share more on this one. Yeah. So I think first, maybe just start off with the definition of AI. I mean, AI is more than one thing. Right? Sure. Like and I say, you you can from kind of the left end of the the spectrum of, like, machine learning and and other capabilities to the right end of more agentic. I'd say we've been doing machine learning for many, many years. In fact, kind of on our platform, you know, if your campfire wants to, you know, get a new, you know, advertising account and you want, you know, a preapproved credit limit that's largely all entirely through machine learning and AI, you get your answer in a few minutes, and, you know, it's it's a pretty effective process and, again, well controlled, etcetera. We've been kind of on the journey to, you know, put kinda more agentic capabilities into place over the last, you know, two years. We've been experimenting a lot. I'd probably say last year was, you know, a year 2025 was a year of experimentation where we were just trying a bunch of different things and seeing what what stuck, and some of it was worthwhile, and a lot of it was good learning, of what not to do and and where the the the it has value and where it doesn't have value. Obviously, one of the open questions we all asked ourselves, was how do you actually control AI, in in a finance ecosystem? And so that was a big focus for 2025 as well and even into 2026, to make sure that we felt confident that, you know, we could put more agent capabilities into production and maintain the integrity of our financials. And so I would say, like, how did we get started? I I think it starts with culture. MetaWave, pretty innovative culture. Yeah. I have a CEO who is, pretty innovative and is, I think believes AI is the thing. So it's probably good, to do to do experimentation, to learn. I also have a lot of partner organizations that are using AI and within Meta. And so if we you know, they're upstream of us. And so if we don't understand what they're doing and understand the the technology, we were at risk, right, of of of at least not understanding our risks and understanding what what could go wrong. And we have a lot of internal tools. One of our advantages, we're releasing a lot of just meta specific tools, and and it's been great just to learn, and it's clearly a huge advantage. I'm I I don't discount that. I know a lot of companies don't have that that capability. But, when you have kind of the culture, the tools, the mindset, you have, you know, peers to learn from, it just created this desire internally to say, well, let's be at the forefront. Let's figure this thing out. Yeah. And huge shout out to Meta, by the way. We actually, use your models internally, and and we've done a lot of post training with some accounting data, and we see they've done really well for accounting workloads. So, congrats on, on everything your team. has done there on the AI front. Yeah. They're getting better too. On getting they're just unbelievable. They're getting. better. So thank you for that. Yeah. Yeah. No. There was, and there was some recent models that just came out as well. In. terms of the controls framework, let's talk a little bit about that because probably one of the hottest topics right now is, like, deterministic versus probabilistic. And Yep. how. do I, you know, I For those that aren't familiar, it's more of, like, human reasoning and in an audit controls environment, human reasoning needs controls where if it's rule based or what's called deterministic, it's like a calculator. It'll give the same answer a million. times. One plus one will, like, always be two, But you ask a LLM, and, occasionally, they might say three. And so then you get into this controls issue and segregation of duty issue, and then auditors that need to, like, certify a workflow have trouble doing that for the occasional three output. How have you taken on the probabilistic nature of models, and what approaches did you do, you know, differently, particularly in a in a SOX environment? Yeah. It's a it's such a good question. The the the phrase I like to use, and I know chat GPT has gotten better at this point, but I've used a phrase, look at at this point, if you ask chat GPT how many r's are in the word strawberry, it'll tell you two. And I said, no. Do we wanna rely on that to moving money? You know? Like like and so, and and, again, back in 2025, and it was really about a year ago, literally a year ago this week, we sat down with kind of I I could see that, you know, look. We we have a team that wants to innovate. We have an innovative culture, but no one really knows, at least I don't know, how would I develop a control environment around AI and finance. And so, it was when I, early in my career, I said I started with national office at at Deloitte. And I remember to this day, the first day I walked into national office, the partner came into me and and said, hey, Aaron. The the you know, you're gonna get a lot of difficult questions, questions that don't you know, haven't been answered before. The odds of you getting it wrong is one divided by the number of people you talk to. And if you sat there, I sat there thinking, boy, you have zero confidence in me. Clearly, like, if I just talk to myself, the odds of me getting the answer wrong is a 100%. Right. But the the point he was making was, look, anytime you're doing something new, anytime it's there's complexity, you need to seek different perspectives. You need to, you know, ask very smart people. You need to bring people together. And and so about it literally a year ago this week, I reached out to some friends at at Google, Walmart, and at ServiceNow. And we said, hey. This is looks like AI is a thing. Looks like, you know, either our, you know, CFOs have been out there talking about, you know, the the the benefit of of AI. We should probably figure out a way that we could deploy AI responsibly. And so I asked them, would you all come to Menlo Park here in California? And I'll promise to feed you, but we're gonna free food. right. Exactly. But we're gonna put you in a room for two days. And the ask is that at the end of two days, we have the start of a framework that we could develop, to deploy AI effectively. The team came the teams came out, so along with the meta team, and, just literally locked themselves in rooms. I mean, it was it was probably in in one of the most impactful things I've been able to watch in my career because you saw very smart people get smarter as they learn from each other, as they invested in each other, as they ask questions and try to figure out answers together. And it was end up being about fifteen, eighteen people in the room. So it wasn't a huge group, but at the end of it, you honestly could not tell who worked for Google, who worked for Meta, who worked for ServiceNow, who worked for literally, they were operating as one team. It was probably one of the most exciting things I've ever I've seen in my my career, and the output was spectacular, from a first step. And what I learned from that was, you know, again, we we know how to apply technology. We've done this before. We've, let's leverage what we know, but then let's also understand what's unique about AI and what maybe we have to do. So we we got a framework together. We then sat down with the all the big four national offices. One of the benefits is you kind of access to kind of all the big four. Sat down with each of the and then this was probably in let's see. We did this in August, September. Maybe early November, we sat with all the big four of 2025. And it was it was and we sat with them all individually, gave them a draft framework. They reviewed it. They came back and said, we disagree with so much of this. And what was funny was. they all disagreed with different things. And what was clear is that there was a lot of learning that needed to be done within the audit profession too. Right? And so we engaged in dialogue with each each of the firms. And what was really neat was, you know, let's say one firm said, hey. We totally disagree with multimodel validation as a reasonable control. Another firm would say, we think multimodel validation is a great, you know, independent control. So you have. these very smart people all taking very different opinions on on, you know, something. And we were able to kinda bridge, bring them together, and try to, you know, resolve. And we saw by February is there was tremendous alignment. And I'd say learning from the big four too, like learning around this where we kinda had a framework, the big four. We're like, yeah. This is this is great. We then took that to the SEC, and we sat down with the SEC staff and said, okay. This is what here's what we're trying to do. The objective is to create a a principles based framework that would allow us to deploy, you know, in production, you know, agentic capabilities, prob more probabilistic capabilities. And we had a couple rounds of comments with the SC. SC was tremendous. Like, I would say, like, I the the feedback was was was really good. They clearly have I just always remind you how smart people are down in Washington, down in the SEC. Had two rounds of just really interactive dialogue. And after the second round, we kind of asked them this was this was back in early June, so it took kind of time to iterate many versions. We asked them, okay. What do you think? They said, we think this is needed right now, you know, in in production. Could you go ahead and publish it? Right. So we did. So we said, sure. We'll publish it. With the idea, we expect this to be somewhat of a living document, meaning anyone who thinks and I I put myself. If you think you know, you probably don't. Like, you you there's just I feel like I'm learning so much every day. And so we said, we'll put this out there with the understanding. It's it's probably right as the technology sits today or close enough, but we expect it to, you know, to need to update it and improve and improve. So we put that out with the hope that we could help our profession to your point of, like, you know, a lot of us are facing that challenge, and we wanted to make it available to the profession of how we make you know, can adopt AI by leveraging what you already know and kind of building off that into what's maybe unique about about AI. We published it under under kinda as part of FEI, which is the group kinda where all the four of us were all members of. We had all the FEI CCR member companies, so 35 of your biggest companies review it also and give input, into it. And, it's behind a paywall there, but, again, I think it's probably pretty reasonable for most people, to get access to. Yeah. And and for those that are curious, before you start searching for it, we're gonna put a link to it in the chat. It's a great piece. I think the big for those listening, you know, maybe we're not gonna go meet with the big four national office, but the big theme here is fear of the unknown. Absent a framework, it's hard to know what is and is not allowed. And so what I. always encourage finance and accounting teams, particularly if you're audited or in the public markets, is draft your own, AI policy and share it with your auditor and proactively do it just like Aaron did before entering it, you know, into the audit environment. And say, here's here's what we're thinking of doing, and here's how we're planning to use AI, and here are things we're gonna we would love to do, and here's things we would not love to do and have that conversation and and preclear. You can all take what Aaron did in your own vein and and proceed. But I think now that there's a clear framework and and and then folks can say you know, Aaron can hand it presumably to your team and say, here's here's how we can operate as a team now. We we have clear guidance because I think that's one of the biggest things. So the comment earlier about AI adoption for, accounting teams in an auditor or a SOX environment, I think just knowing knowing the lines is one of the most important things to adoption. So, one one follow-up there, you know, why, Yeah. why build, you know, your own guidelines on the existing controls? I think, anything you wanna add on before we move on? Yeah. I think it's you know, that's a great question. I didn't we didn't go into the exercise expecting to do that. We thought, oh, AI is this brand new technology. It requires a way new way of thinking, etcetera. But as you started we started going down the path. We started thinking, well, don't you just do a risk assessment? Understand where it could go wrong, don't you? And we started, like, using and we're like, wow. Like, we actually the the learning for us was we actually somewhat know the framework by which we go which should go at this. And and so we started applying through those two days, like, existing k building. And I don't think it was we were very cautious of making sure we don't just rely on what we know, right, and and building off that. But we learned that the actual, you know, framework of do you understand the technology? Do you understand the risk? Do you understand, you know, the the data, etcetera? Like, it started to become, okay. We can apply the guidance. We can apply the guidance. And then when we finally sit down with the SEC and said, hey. We actually think because it was funny. Even the SEC in December, when we told them we're doing this and at some point we're gonna come to you, they asked the question, do you think our existing SEC guidance is sufficient to and my answer at the time was, we haven't found anything that's that's needed yet. Sure. And the SEC's reaction was basic, and I felt kinda bad because their action was, well, then maybe you just haven't looked hard enough. Was that they're they're like, well, you're probably wrong. And I'm like, oh, boy. Did I just look stupid in front of the SEC by saying, like, I I I think the guidance can be applied, you know, against uniqueness here and there. But and then by the time we got to the conversations in, you know, May and June, like, they were like, yeah. No. We think you've applied our guidance. You've we don't think there's a new guidance that we need to issue. So it was a little touch and go there. I thought I had, you know, kinda made a mistake, but I but I think that, you know, that's been helpful because I do think and that's probably the encouragement I would ask I would just say for everyone on this call. Like, don't underestimate the core skills that you have, in especially if you're an internal control minded person and others because you have the absolute framework, skeleton scaffolding, whatever you wanna use to be absolutely special in this area. It's what's needed. And I'm I know we'll probably talk about the profession, but I I couldn't. be more excited for the profession given given the benefit this core skill sets we bring to the table. that is needed around AI. And one just side note, in a prior life, I had to present to the SEC for a couple hours on a technical accounting piece at a public company on, we had a payments product and gross versus net, our own auditor. It went to corporate. They said, we're not sure. So we went to the SEC and they were truly amazing to work with through that whole experience. Yeah. In in terms of, like, you mentioned it's a living, breathing document. So I think the last question on this before we move on is where do you think traditional SOX is, like, not ready, or not work for for AI? Or just, like, where are you you know, anything you wanna share and, like, where you think the document's going next? Yeah. Great question. I don't yet know where the document's going because I think we're still learning and applying it and trying to figure out what maybe is missing, what, is confusing, etcetera. Sure. So love feedback, by the way, because I suspect maybe in a year from now, we'll probably be saying what have we learned and what needs to be updated. Your your question on SOX is an interesting one. And in fact, one of the other questions we we got asked was, do you think the PCOB rules, on, you know, audit guidelines, are up to speed to in a in an AI enabled world. And, you know, the the answer at least my answer is on this one is maybe not yet. And where I'm going is, you may and, again, I'm I'm old, so I go back to the old days. Like, before IT controls, before SOX, we would do things like audit outcomes, like, you know, accounts payable. We would do subsequent, you know, disbursements testing. And and, again, the ultimate goal in an audit is, you know, are the financials accurate? Right? Sure. And sometimes I feel like as a profession, we've lost that perspective. We're we're more auditing is the process accurate and not is the outcome accurate. And I think AI is forcing us in in I was at the PCOB down in in December, and we've raised this with with the PCOB, and they nodded and understood. I said, look. I think you need to understand, and we may have to go back to the future a little bit. We may need to go back to how do we, as a profession, get comfortable with not auditing the ones and zeros and the the audit trail per se, but we audit the outcomes of the technology. And I don't think we're there as a profession, but I'd say that's probably the one area in a SOX world that we need to think about. And I I do think there's there's a lot of could be a lot of value in our profession if we can get. out of some of the, you know, archaic, you know I mean, most of us probably have, you know, your basic deficiencies that you you have around some, you know, IT control thing that, you know, is it it's just gotten out of hand a little bit. And if we can get back to, what's our core objective is accurate financial accurate. outcomes and be able to do that. I think that's probably where the Sox world will will need to evolve. Yeah. I think I think that's well said. Let's talk a little bit about how your team's using AI internally. And so, obviously, it's a very large accounting team given just the size of the business, very complex, you know, operating globally and the, the bigger theme, like the larger teams, I talked to a lot of finance and accounting leaders and it's how do we make sure the whole team is moving into this era? And, you know, I think where there was, like, on prem to cloud or, like, you know, my my old boss used to tell me he was, like, using graph paper to do his charting. And I was like, yes. I get it. And so I think there's always a big there's always a big transition. You know, now that you've got the guardrails, you've got the framework in place, how are you most effective effectively using AI and accounting at Meta today? Yeah. So let me start with so, obviously, individual productivity, people are using it, you know, for presentations, all this stuff. So course. I'll skip through that. Like, that, I think, Writing emails. and performance reviews and Exactly. Exactly. Although I I've tried not to do that. I feel like doing a performance view via AI is just so impersonal. So agree. maybe that's my age showing up. But, so we're in in production, and, again, I'm gonna go back to we've been last year is year of experimentation. We've been building a lot of the core, capabilities. So back in 2024, we kinda saw this coming, and we recreated a whole data team and redid our data strategy with around certain pillars of data knowing that data is gonna be obviously critical to the use of AI. And I would probably say one of the the just core things is is we've moved from what I'm gonna call a data strategy to a data ecosystem. So if I could just leave leave one thing with don't underestimate the complexity of a data environment to enable AI, in truly what what you wanna do. And what I what I mean by this, Yeah. and I don't mean, John, to go off on topic on on how we're using it. I'll get into that. But I I feel it's so critical. And what I've learned is that in traditionally, when you think of data and finance, you think of transaction data. Right? Your traditional debits and credits, employee data, etcetera. Like, you're you're again, or, you know, payroll data, like, you know, your basic transactional records. As you get into the world of AI, you start to add things like, you know, nonfinancial data. So this would be forecasting and budgeting and and kind of more employee data, you know, beyond just your traditional kinda cash aspects of it. And and that requires a whole different strategy if you're gonna bring it all together, and then you add to that contextual data. So if you wanna use it for things like, you know, hey. If, you know, there's a geopolitical event, how does that impact my supply chain, etcetera? Like, you're talking about all that contextual data that you need to bring into your ecosystem. Then you add on to that, Like, if you're gonna do truly control testing of AI, you're gonna have to have certain controlled datasets to test your AI. So now you have what I'm gonna call the internal control data, the data. that you're gonna maintain and monitor to make sure your your AI is continue working. So you've gone from the first pillar of just having transactional data to now having nonfinancial data to having contextual data to having internal control data, and it broadens the whole ecosystem. So I would say first and foremost, I would say it's and I've what I've learned is every attempt to try to do something in AI or something around AI within finance hits a limitation on data. It just its usefulness gets limited. You have to have to have to focus on on getting that data and having it controlled, making sure you have that right access rights and the the right controls over that, you know, contextual data and the non the nonfinancial data and the testing data and everything. So that's number one. Like so that was a learning too in 2025 even though we started this in 2024. In 2025, we were starting to do things like, well, how do I create an AI CAO? And how do I, like, Sure. you know, try and we're just running that. Well, the data was inaccurate or it was incomplete, and, therefore, the AI was trying to guess, without the right data. And so we just found it was extremely limited on what we could do. So what we then went to is we said, okay. Let's look at the data we have and what in control, then how do we do that? So, obviously, we have GL data. We have transactional data. And so for one of the first things we built was like a flux tool. So how do you you know, normal flux, you know, in the old days old days, Okay. you know, you you have, you know, you have GL data, you have transaction database, you have a you go you're trying to pull and say, why did, you know, the balance go from a 100 to a 150? Right? Whatever you're you're doing. in a balance sheet flux, and you're going to a bunch of different datasets. Well, generally, it's the same datasets. It's the same and if you have the data hub, you can train the AI to actually write the flux analysis for you, to go to the data, analyze the data, and then provide a flux analysis. We're now in our fourth quarter of of doing that live. What I would say is the first quarter, was was good. It was, and and we we measured it in terms of how accurate was it, how did it need to be, you know, changed, and we started. And then we just expanded it as and then the AI learned from itself. And now, like, last quarter so third quarter doing it, I'd say it was pretty, pretty good. In fact, we we have a a seven day close, which isn't the the fastest close, but we moved to a six day close just from that, just from automating Flux and and going through kind of all the analysis. And you would imagine then when you know kind of that core data, what happens, and it kinda is a nice framework to to add other capabilities. A big portion of my team also shared services, and we support our revenue, operations. And so we have, agentic capabilities supporting our customers. If they want, hey. Can you send me my invoice? What was this transaction for? Those kind of things. So we're built we built kinda AI. We're still in pilot on those. They've been very effective, but we're, obviously, from a control perspective where, you know, anytime you're dealing with revenue and customers, you wanna make sure you're you're exactly right. And and, again, I know I said it earlier, but I'm gonna say it again. Like, it is taking probably quarters longer to deploy something that you trust than what I would have expected. Okay. I think we've been impilot on the revenue for nine months now. Just just for example, because we're just continually modifying and changing, and we're seeing it. Oh, shouldn't have done that. Why did you do that? Let's you know? And and just making sure it's well controlled. And I the thing the phrase I use my team is we earn the right to go to human on the loop or her you earn the right. And so at each step, we earn the right. And, again, it comes back to your core values of integrity and trust. And if it's not ready, it's not ready. So, but those are maybe a couple examples of where we're using it today. I think they've been pretty impactful. That's great. I think Flux is a great one, and it's a great place for everybody to start because, obviously, you can review the Flux, you can confirm everything's accurate. It's not incredibly high volume, so you can review all of them, but it gives you great ideas. When I was at Adobe in the finance team, I would write some of the Flux commentary that would roll up into the earnings call. And we were looking at dashboards, and we were obviously, like, doing our best. But AI can read every deal done, all trends. It can look at cash in, cash out, and find, you know, oh, looks like we saw a margin compression because of x, you know, and, like, things that I just physically would not be able to comprehend as a human writing those. And so you can. come up with some, like, novel ideas by just looking through larger volumes of data, but you can still go in and tie it out and and review it on your own. Is that where you're seeing the the largest impact on the on the flux and and hopefully soon on the revenue? Yeah. We're that, I think we're also seeing it. We're we're building what we're calling our second brain on the the general ledger and our controllership. Oh, cool. And, really, the idea there, it sounds really fancy, but the idea is it's you know, you're basically getting rid of desk procedures. I mean, they're just you're you're you're kind of educating, you know, kind of the the the core system to understand what needs to be done when, and we're we're we're automating, you know, a lot of that activity, but it also helps as as we have attrition or we have transit people transitioning jobs. They come into, you know, a more well documented, understanding of what has to happen in in that role. So I'd say that and just core operations. You know? I think, you know, we're we're also we leverage a lot of outsource, you know, outsourcing, and a lot of that is manual. And and that's been obviously a a key opportunity for us just for cost savings to bring some of that manual activity and and just reduce it. So we're just a little bit more efficient productive. Yeah. And just to just for the group, for the next question, like, what's the size of your team today or anything you wanna share? Just so folks can kinda understand the magnitude. Yeah. So as a finance team, we're about a thousand. As an accounting kind of my team is, is roughly about 550, plus, outsourcing of equivalent of about I think we're about six fifty now. So you can kinda 650, Okay. like, Yeah. human being than outsourcing plus 550. So, about a thousand. And but I I suspect, we'll we'll get more efficient, especially on the outsourcing side, Sure. given it some manual activity as you would imagine. Yeah. And so let let's for the five to five fifty ish on your team, how do you. think about AI adoption? You know, I I think, like, some of us are saying, how do I get AI adoption in a team of five? You're you're you're clearly on a magnitudes of that. How did you approach it approach it? Anything you wanna share, particularly for folks listening that have a larger accounting team? How do we, you know, they say, like, speedboat versus cruise ship. It's certainly a larger team to to steer. How do you help them move into the AI era? Yeah. It's it's a it's a great question. And and the way I I go back to like, I I was able to hear listen to a psychologist many years ago, not not my own, but someone about just organizational change. And and they said, you need to understand that the in any organization, there'll be about 20% of your people that will just be gung ho to whatever's new. Right? Whatever's new, they're just gonna dive in and love it and and just go. Right? There's gonna be 10% that are gonna be like, you know what? It's not for me. I don't think I can do it. I'm close to retirement. I don't wanna learn this new thing, Whatever. And then you have your middle 70%. And and, again, those are obviously directional, but it's pretty pretty accurate. My my goal, and it and it's been in any change, is to focus on the 70%. I'm not gonna focus on the 10 that that have already agreed they're they're not in it, and I'm not gonna focus necessarily on the 20 who don't need any motivation to to try something new. And so it's how do I uplift kind of the the 70%. And one of the things I saw kinda early on was and and, again, Met is very we we hire great people. We have tremendous talent. Arguably, some of the the best talent I've ever worked with in my career. Well, not arguably. Absolutely the best I've ever worked with in my career. And what I saw and then you have the engineering side who's, you know, doing things, like and showing it off. Like, you're like, oh my gosh. Like, I I have no idea what the heck is going on. And what I saw happening was you had a very small group of people running ahead, and you saw this huge group of people going, oh my gosh. I'm falling behind. I'm no good. And I saw, like, I had started to worry me. Like, these are incredible people, talented people, and and I'm sure, like like and I was a part of that group personally. Like, I I felt, like, so dumb, so behind, so like, oh my gosh. Am I gonna be able to learn this? And I was like, you know, how if if I'm feeling that way and and my people on my staff were sharing that, and I like, what's the organization feeling like? And I felt like, you know, if I don't do something to bring up people like me, the the, you know, the the people that feel behind. And, again, when you're realistically, you look outside of Meadow, we probably weren't behind, but we felt very insecure, very like, oh my gosh. Can I really do this? And I'm sure many of you are probably feeling that way too. Like and I think it's normal. Like and so what we did is is I said, look. Can we just define what AI fluency means for a finance professional? Because I feel like when you're comparing yourself to, you know, some engineer, someone who, you know, has an engineering background, computer science, whatever, and they're apply like, you're just gonna have the wrong comparison. Yep. So the first step was define AI fluency. And and so what we did is we defined it in four main pillars. Do you understand the tools? Do you understand, the mindset required to to approach AI and the process, etcetera? Third is is do you, do you understand controls? What we just talked about the framework. Do you understand risk and responsibility? And then the fourth is do you understand how to use data? How to design it, how to make sure when you're designing something? And and then we defined in each of those four, we we created content. How do we build up? How do we then define the aspects of toolkits? Like, what tools do you actually need to know? What we're finding is that as a as a and I like to say that the model we use is is my sister is a is a marine, and she you know, every marine has to be a martial arts expert by far. Every marine goes for martial arts. They call it level zero martial arts, so everyone has to be level zero. Now certain marines go level one, level two, level three, and I brag that my sister is the only woman in marine corps history to be level three, Wow. and she she has a reconstructed nose to to prove it. But, but the idea was how do we build that baseline we call level one that every, you know, AI every finance professional is a level one AI fluency. That's what we're doing right now. And so we created this async learning, and then what we decided is we're gonna take three days. Just we're gonna every person in MetaFinance is gonna go through a three day after you've done the async, go through a three day in person training to build up your AI fluency. We've piloted we've done two. We're I have our third one actually right below me, going on right now in. MPK or or in Menlo Park. And, and the idea is every single one of the five fifty will go through, not only async, but then the the the in person learning. We're gonna invest. We're gonna travel. We're gonna do those things. We're gonna invest in our people. The feedback so far has been tremendous, and and and it's been exactly what I was hoping, which is, that I got an email from someone recently just saying, basically, I was about ready to quit because I felt I was falling behind. And it's phenomenal. And and their performance was going down. You saw the performance because they just didn't have the self confidence. And then you see them, wait a sec. I can learn this. I can build it. I I I know what's happening, and it's been so much fun to see, like, the performance change. And they're like, no. I can do this. I can learn. I can grow. And then and then, honestly, we've had a couple people leave the company, that were very AI, you know, fluent, and they went on to phenomenal jobs. They doubled their salary, both of them. They did and I'm like, that's success too. So, like, I feel like as we, you know, think about how do we invest, it's investing in people, giving them just a little bit to get forward, and then celebrating the things that they learn through it. So sorry. I know it went a little long. I bet you could see. this is, like, amazing. the the most important thing I could ever leave is you can learn it. And if you feel like you're like me where you you're at that point where can I really do it, you just keep reminding yourself you can, and there's times you've done it before? Yeah. I mean, look, there's the there's the selfish thing of, like, I want my team to move into the AI era just because it's good for us as a team. And then there's, like, the humane, you know, moral thing to do, which is like, let's move them to the area for their own careers and professions and, livelihood. And and maybe some will go on and take new jobs and some will stay. But ultimately, Yep. I think that's a great call out. And I think the other big theme here was, like, frameworks and education. And I don't know if you called it a boot camp given the marine corps and the family, Yeah. but, We didn't. But yeah. moving moving people along, and and writing it all down. I I think. that's something. I've learned actually is, like, you need to write everything down as the as the team grows. because you talking about it and then repetition is the other big one as well. We're gonna wrap up. with, just a couple questions on, like, the road ahead and and your vision for accounting. But before I do, we've got the q and a coming up right after that. So please start to put some questions in there. We would love to hear from you. We can preload them now. So start to put them in the chat as, Yep. I'd love to hear what, Aaron has to say from you all. But, last question or two, you told me before, you know, faster close that you said you're at a seven day. I talked to some folks and say, I'm already at a two day, one and a half day. Who cares? So I think for. you, like, what is the end game if it's not a faster close? What's what what's your vision? Yeah. It's a it's a really good question. So, we we kinda look at, you know, what's the value of three days or four days or five days for us? And I have the benefit of being a founder led organization who thinks long term, who isn't, you know, pinging me on, you know, day minus, you know, the the day before close asking what the numbers are. So I I don't have that that pressure, which is great. For me, I'd say I I'll I'll say tactical and then maybe aspirational. Tactical is I'd love to get to more of a continuous close. Okay. I'd love to be able to to do continuous control monitoring, be able to identify, you know, issues as they arise and not wait until, you know, the quarter end processes, etcetera. And, again, we're we're already doing that in certain areas, and I I just wanna expand it out a a bit more. I think that the other questions were, like, John, you and Ramp are are thinking through a lot of this too is, like, how do you just make close a nonevent? How do you rethink like, we're all geared of, oh, every month, you close the books or every quarter, you close the books and you're well, how do we just get to a point where an event occurs and you're and it just goes all the way through, not only through your financial systems, but it impacts your forecasting and your so I think the future is is kind of more dynamic forecasting, more dynamic, you know, close, etcetera. And I think kind of breaking the what I'm they call the HR barriers between, like, organizations. And we're just all organized so much by our org charts, and how do you kinda bust through that a little bit, and go across. So that that's kinda one of my, probably, my end goal. And by the way, I think we will close faster. I just don't want that to be my goal, if that makes. sense. Yeah. I mean, on a long term vision, I think you're clearly thinking beyond that. And to that thanks for the shout out. Yeah. We have moved, internally, obviously, we're on Campfire. We have moved to a continuous close, but also it's allowed for continuous forecasting. And. and with that is, like, we've got agents that are in the Campfire environment, and they're out, like, looking into other systems like a ramp or like a a CRM, and they're grabbing kind of forward looking non accounting objects like a pipeline or taking other things and rolling that forecast for us. And so it's been amazing to kinda see what AI is able to do, reducing re materiality thresholds, tightening variances, more granular than we would normally do. Yeah. So I completely agree with that that vision. In terms of, you know, you've talked a little bit about, the Excel analogy in the past when we chatted. You know, people feared Excel went into accounting. You know, would accounting software and accountants, you know, as a t t TSheet, I think, before, TChart. You know, tell us a little bit about how you think about the profession. Yeah. Well, John, I'm I'm old enough to remember when I started auditing. I I was the auditor that carried the computer to the audit because I was young and my back was strong. So I How much, in a way? It had to be a heavy. one back then. I I remember it it it weighed enough that you needed to have the big audit bag on the other side to offset you. Right? So you had nice balance, as you're, walking it Aaron was. walking like this? Exactly. You have your shoulder, like, down. And and it was funny because I just started accounting. You know, Excel came out, and there was a headline in The Wall Street Journal. It said, you know, accountants are gonna only work thirty five hours a week. The job of accountants are gone. Who needs accountants anymore? And I remember my parents, and I'm I'm the first one in my family who go to college, and my parents kinda called me like, did you choose the wrong profession? Like, should you, like, choose something else? I mean, what what are you doing? And I'm sure a lot of kids coming out of college with the county are getting that same question from their parents, right, that I got back in the day. And and I think we all know kinda what happened. Right? I don't think accounting jobs went away. In fact, it created a new, you know, a new thing called data analysis and new capabilities, and and and we started using technology. And then RPA comes out, and we we've seen just technology, you know, innovate ERMs, and, like, you know, we're gonna see what Campfire does to the whole system. And, you know, you're gonna just see, like, continuous innovation. And the question, though, is, like, what's the role of accounting? And and the reality is our role is is we verify truth. And what that role is more important now than ever. And. so I do think whether you you you know, I know there's a professor at Arizona State University called, you know, Eldar Maximoff, and he's done a a paper on the verification imperative, which I think is is spectacular. But he's you know, again, goes back to Excel example. He goes to Javon's paradox, which is again and but he also talks about, you know, AI may may change, but the the fraud, you know, the risk of fraud, the risk of error, there's it does not change. In fact, it only accelerates in an AI world to a certain extent. So I do think it's gonna be imperative for us, as a profession to lean into that, to learn the skills. We. have the baseline, to to be able to Amazing. Well, we got some questions starting to come in, Aaron. do that. So. and totally understand you're operating in in the public company environment, and so answer answer what you can on these. Yep. The first one on the Flux example caught some folks' attention. When you said tool, was there a separate internal tool or teaching an existing LLM or anything you're comfortable sharing on how you, how you achieve that? Yeah. So we, leveraged so, we started off doing competing models. So we actually have access to every model. So we have we had, you know, Google's models. We've had chat g p t. We have anthropics. We have our own. And so we ended up, competing the models. I think we used OpenAI's, model to begin with. But, obviously, given our own, we we then we're. kinda waiting for metas to catch up. Metas did catch up. We've transitioned over to meta you know, our our own models at this point. But, we used so we we we did competing models. I do think it's important some models are better at certain things than others. And, but, yeah, we use, kind of our LLM right now, and then we've kind of baked that into our kind of our internal tools. So we have kind of our internal engineering team. It enables certain tools, and we just use those. Kinda within we have a closed platform, so we've made it kinda part of that closed platform, which is a an internally developed platform. Very cool. Yeah. And we're we're similar. We have our own models and we commercial and then we route based on performance and also by task. And we're now. taking a look at, like, larger tasks and breaking them up to various models depending on, cost, performance, accuracies. Yep. It's a great time to be a a model customer. Another one coming through. Can you share an insight on AI tooling or models your team uses internal kind of along the similar lines, but I, I think the more important piece here is like, how do you set user spend limits? I know on the open source side, maybe it's a little different on your own models. Yeah. Like, on the commercial models, how do you think about spend limits, if any? Yeah. It's a it's a good question. So I think where I am right now is we do not have spend limits. Okay. And we're really fortunate that that Meta has has prioritized right now. We think there's so much value in learning that we're not gonna limit spend. That being said, I've kinda worked with the team to say, look. We need to understand TCO of of every you know, total cost of ownership of everything we're putting into practice. Like, you know, I don't wanna be, you know, spending a million dollars in tokens to save a $50,000, you know, outsource cost. It just and I feel like it's but it's intentional right now to say, let's learn, let's experiment, and then let's refine. We haven't got there yet, but we're we're gonna start doing that. And I've I'm asking a lot of questions on my FP and A team and, around this, and they're starting to build the capabilities to to be able to manage token use or at least report on it. I I we don't do, like, we don't measure people in finance and how many tokens you're using, so I'm not worried about, like, token maxing and, you know, Okay. just international spend. So I but I you know, it's something I watch. And right now, it's it's a reasonable cost, so I don't I don't feel like an urgency to to change it. Yeah. Amazing. And I know you briefly touched on it earlier for the piece that you coauthored. You know, anything like, if someone weren't to download it, any, like, kinda quick summary or sound bites for the group, that you want like, what's the overall, like, punch line that you'd wanna share for folks to take home with? Yeah. I think that the the punchline is you're gonna recognize a lot of the wording and terminology by by design. Sure. It it will be something that you can you know, 75% of it will be something that you're like, I already know this, to. a certain extent, but then the 25 will be, oh, yeah. This is okay. This is a little bit unique to AI, and this is what we need to do. And, the other thing I'd say the punchline of is we've actually built an an AI skill within kind of our meta, we'd call it MetaMate, which is our own internal, AI tool, that is applicable to ever anyone to download. Whereas they're applying something new, they can then take we call it Sage. They could take Sage and apply it to and have it review kind of their process and give them feedback on, hey. You missed the step, or you need to inventory this, or you need, you know, part of this skills you know, part of this AI would be to do this. So we we've actually then kind of created this AI skill that, you know, as you're doing more citizen development, you you can then apply the the the skill against and give you advice on how how you actually build it with controls in mind. Awesome. That's amazing. Well, we got one final fun one, and then we're gonna call it. You mentioned them at the beginning. So what are your thoughts on the Huskers this fall? Is was one of the questions in the chat. What a great question. The most critical enough time. of the day, Aaron. It is the most critical question. I think as a as a lifelong Husker fan, I've. learned that hope is what kills you, I think, is the old Ted Lasso. Right? It's the hope that kills you. I've I've now got to the point where I'm just trying not to be hopeful. But, again, I will I will learn, going forward that I was probably more hopeful than I thought. But. they have to they have to be, more competitive this year. I. just they'll make a bowl, but I don't know if they'll go beyond that. It's just. a different environment. Right? College football. I'm sure the NIL is reaching out. I'm sure they could use some helps though. Exactly. But, well, look, just wanna thank everybody for joining. Aaron, huge, huge thank you. I know you gotta be incredibly busy over at Meta. Amazing work on, like, helping advance profession. I think just wanna thank you for the the paper. and all the and continuing to just think about what's next for accounting in the in the public markets and just for all of us. And, if you wanna move to Aaron's vision on the continuous close, check out campfire.ai, book a demo. We would love to chat with you. And, Aaron, I just wanna thank thank you again for the time. So hope everybody enjoyed that, and, have a great day. Thanks, Sean.