Latest / The Accounting Podcast / The Death of the Financial Close: Aaron Harris on AI's Accounting Revolution

Transcript

The Death of the Financial Close: Aaron Harris on AI's Accounting Revolution · The Accounting Podcast

• Attention: This is a machine-generated transcript. As such, there may be spelling, grammar, and accuracy errors throughout. Thank you for your understanding! Blake Oliver: [00:00:11] Hello everyone, and welcome back to the Accounting Podcast. I'm Blake Oliver. David Leary: [00:00:15] And I'm David Leary. Blake Oliver: [00:00:16] And we are at Sage Future in Atlanta, Georgia, joined once again by Aaron Harris, CTO of Sage. Aaron, welcome to the show. Aaron Harris: [00:00:24] Hey guys. Thanks for having me, and good to see you again. Blake Oliver: [00:00:26] You know, you've been on our show several times, going all the way back to 2019, and I went back and listened to some of those episodes. And believe it or not, in the very first episode we were talking about automating the financial close, eliminating the financial close. Yes. Aaron Harris: [00:00:44] So there's evidence that I have been talking about this for a while. Blake Oliver: [00:00:48] Very consistently, actually. Um, and you were talking about it before we had generative AI. Aaron Harris: [00:00:53] Oh, yes. Blake Oliver: [00:00:53] Yeah. Because that's been, what, three, four years? Aaron Harris: [00:00:56] Yep. Three. So. Blake Oliver: [00:00:57] Um, I would love to get your take on going back to 2019, when you first started talking about this stuff. Or maybe it was before that, even up to now, like the journey that we have been on when it comes to AI in accounting and finance. Aaron Harris: [00:01:13] Yeah. Well, I'm obviously very happy to talk about that. Um, yeah. There's, uh, we like to think of it in waves. Um, so, you know, we went through this wave and it's still a very important wave, the AI task wave. And this is fundamentally where we're developing AI models that are very specifically targeted at one thing. Um, and in fact, you'll often need lots of models to work together to do the one thing. So like reading an invoice or classifying a transaction or looking for anomalies. And what this allowed us to do was to automate a lot of these processes, like reading and categorizing an invoice, but we had to do it in a very controlled way. You know, this I was was very scripted, if you will. Um, it followed a predefined path to get to get the job done. And, you know, humans were required kind of at every step of the way to sort of move things from one step to the next. So, you know, the big problem is that you can't really interact with this AI. Um, and because these are sort of very narrowly defined models, they can't do a lot very flexibly. So large language models then open up this era of generative AI where, you know, from the way I look at it, the way I like to talk about it, the two big things are now you can interact with the AI, um, and it's those underlying capabilities allowing that interaction that allow the AI to work more flexibly. The real breakthrough. And that's the generative wave, right? The real breakthrough comes with Agentic AI, where we're now equipping these large language models. Aaron Harris: [00:02:54] We're putting them into software systems where they can access tools and they can interact with with other systems. And we're putting brains in them that that don't just predict the next word in a, in a sentence. They sort of think through how to plan something start to finish and that they execute on that. So, you know, we've come a really long way. I told our customers this morning that prior to 2019, when I first started talking about this 2 to 3 weeks to get the books closed, our last close, the book survey, was a week. Um, just recently we had some customers at our new office here, and they were telling me 2 to 3 days. So we've we've gotten really, really close. Um, you know, and I don't think this is, you know, asymptotic, right. Like, it's not like going to go slower and slower and slower and never actually get to zero. I think, by the way, I used to believe that. I actually think now there there are some breakthroughs that we're going to reach a point where businesses say, you know what, we're just not going to operate this way anymore, right? It is so easy to ensure we've got accurate, real time information that we're not going to worry about a close right when we need to get that accurate, real time information, we'll click a button and everything will be ready and we'll go continuous close. Blake Oliver: [00:04:12] That's the. Aaron Harris: [00:04:13] So here's the thing. Yes. And I know that's like I know that's the vernacular right. But let's take a step back. Why the hell do I need a close. Like, isn't that kind of an archaic concept like this idea? Like, I'm locking up the books, literally, that nobody can access them anymore, and so that the data is like, you know, memorialized forever. That's like ancient. Yeah. So why don't we just get rid of the thing? Like, let's not talk about a close anymore. Let's just, you know, reimagine. We're in this world now where data is always up to date. It's always reliable. Uh, and I don't have to do this whole close thing. Blake Oliver: [00:04:53] No more financial close. Not even a continuous close. It's just. Aaron Harris: [00:04:57] I want it gone. Blake Oliver: [00:04:58] What do you call it instead? I mean, you don't replace it. Aaron Harris: [00:05:01] You try to forget about it, you bury it. I mean, you don't talk about it anymore. Blake Oliver: [00:05:06] All right, that's where we're headed. We're not quite there yet. David Leary: [00:05:08] So now we know what he meant when he said he wants to eliminate the financial close. Just like nobody's going to say it anymore. Aaron Harris: [00:05:13] I pick my words carefully. Yeah, right. I don't want to continuous close. Don't get me wrong. Blake Oliver: [00:05:19] Yeah. Aaron Harris: [00:05:20] I know that progress towards a continuous close is like what? We need to get rid of it. But I'm trying to be a little provocative here. Blake Oliver: [00:05:27] Okay, okay. Let me challenge you on this then. So a big part of the financial close is work papers. We have to produce financial statements based on a period of time, a quarter, a year, a month, whatever. So that's why the close is aligned with those periods, right? We close on a monthly basis. If we want to produce monthly financial statements. And then we have to have all the work papers as accountants to support What the work we did to prove it out for the auditors. Internal and external. Right. So what happens to all of that if you don't have a closing? Aaron Harris: [00:05:59] We get rid of all that, too. Blake Oliver: [00:06:01] Okay. Aaron Harris: [00:06:02] No. So when I talk about this, you know the bigger picture, right? You know what's what's holding back the accounting industry? It's all these cycles. It's not just the close, right? It's the annual audit. It's, you know, the cycle of filing taxes on a quarterly basis. All of these things need to be continuous. Um, and we've got to be connected with our ecosystem and our stakeholders in such a way that there isn't some periodicity to when I can issue reliable information. So and this isn't like pie in the sky thinking. So, you know, it's a really big movement in Europe now to digitize a lot of the compliance activity. So you know, there are countries where when you issue an invoice to a customer, you're also issuing it to the government. Blake Oliver: [00:06:46] Yes. Aaron Harris: [00:06:46] Right. You know, there are countries where your general ledger is continuously synced up to the government, right? So fundamentally, when it comes time to file the taxes, you're just signing something, right? Because they already know what your taxes are. Blake Oliver: [00:06:59] So I knew about the invoices like the in Brazil, for instance, every invoice has to be filed with the government. They have a copy and then every payment can be linked back to that. But what countries are like and what countries does the government have access to the GL. Aaron Harris: [00:07:15] So United Kingdom right. So we call it making tax digital right. So fundamentally what happens. And no it's not continuous. Now they they they sort of enforce a cycle to this. I think it's quarterly. But your general ledger gets uploaded to a government server. And it's got to be coded to the taxonomy that, that, you know, they they indicate I mean huge opportunity for us with AI, right. So we, you know, we use AI to do all that coding for them before they load it up. We look for the outliers and whatnot. But I guess the point that I'm saying is, um, these things are not. There's lots of incentives beyond just. We hate them to move there. Right? They're they're better for everyone. And, you know, so you could look at e-invoicing or making tax digital and you could say, well, that's government being really super intrusive and okay. Right. Like as Americans, we probably react a little bit more than others. Blake Oliver: [00:08:13] A little hard to imagine us doing it here. Aaron Harris: [00:08:15] But what it forces you to do as a business, right, is to actually digitize, right? You know, like you have to take those steps now to modernize the way you do your accounting. And that's that's why we, you know, that's why we appreciate it, you know, and it also, you know, it's pretty good about stamping out fraud. Blake Oliver: [00:08:32] So instead of producing financial statements on a periodic basis and auditors going in and auditing the quarterly financials or the annual financials, they just have what access to the system so they can go in and audit continuously or whenever they want or. Aaron Harris: [00:08:46] So like I'm being really futuristic. Blake Oliver: [00:08:49] Here. Paint us a picture of that. Aaron Harris: [00:08:51] Right. So. So my vision for continuous auditing is that. So first of all, the auditors are going to make a lot more money than they've been making. Right. And the reason for that is they're not going to provide, you know, annual assurance. It's going to be continuous assurance. And they will go in and they will set up technology to continuously audit the books. Right. And you'll have this agreement with them. And you know, they will. And the great thing about it is that this technology will flag when there's a potential exception right before it actually gets booked and it turns into a real issue. And so auditors are going to be working with their clients on a continuous basis to constantly correct. Right. And make sure that things are on track. And if you think about the value to our markets, right. The value to our markets of having way more confidence in the financials. It's it's astronomical. Right? Um, so I know I'm being provocative. I really, really am. But I think this is this is the kind of thinking we need to really, really transform the industry away from one that does sampling and reports on cycles and trust comes on a, you know, a periodic basis. Blake Oliver: [00:10:07] We're already starting to see a bit of what you're talking about in the product announcements from like an internal audit standpoint, surfacing issues, uh, in like manufacturing logistics. You've got something coming for that if it's not already here. Yep. Could you. Aaron Harris: [00:10:23] Yeah. Blake Oliver: [00:10:24] We do talk about that. Aaron Harris: [00:10:25] Yeah. I mean, this was so when we first built the AI team at at Sage. Um, actually, Steve asked me to to build the team before he asked me to to take on the CTO job. Um, at that time, the customers weren't asking for AI. Uh, our product teams weren't asking us to build Built I. So we had to sort of work on conviction and work on principles and vision. And so what we decided was the first thing we're going to build is outlier detection. And the reason why we're going to do that is that, you know, when we talk to finance teams and CFOs, the thing that comes through loud and clear is that, like they need to be trusted. The thing they care about the most is that their books are accurate. And so building outlier detection for us, sort of the way it's manifested to a customer is this capability in the product. But what it did below the surface, as was it, required us to build this massive infrastructure for automating machine learning and, you know, automating model operations and deployment and controls and all of that. Aaron Harris: [00:11:34] One of the big things that it required us to do was to build the ability to automate this on a company by company basis, because an outlier for company A is not the same as an outlier for for for company B. And so this laid the foundation for us to now take outliers out to other parts of the business. Right. We started in the general ledger where we've got it now in accounts payable. Um, as you saw, we've got it now in supply chain. Uh, we've got this really cool thing we've done in construction with, with estimating. So when you go to sort of collect bids from subcontractors, we've got AI now that will tell you. Well, here's who we think we should go get bids from based on history. And actually, are those bids aligned with where we think they should come in for this project. Right. So just building this core competency around outlier detection, we thought was critical to be one of the first things we did because we wanted to signal to our customers, like, we understand that trust is really, really important to you. Blake Oliver: [00:12:36] So give me some more examples of outlier detection. What does it look like in accounts payable? Aaron Harris: [00:12:41] So in accounts payable I mean there's there's there's a bunch of things that that we can do there. The, the the thing that we've tackled first is that we have models that we've developed sort of where they're trained across the customer base for commonly used vendors. Right. So if there's a vendor that we see 100 times across the customer base and we have a patent on this so the competition can't copy it. Um, we've got this way of automatically graduating to a fine tuned model just for that vendor. And it does a number of things that essentially establish a fingerprint. Well, when we get that fingerprint, we know we can run the fine tuned model. That's really, really, really accurate. But if we have something that claims, you know, to to to be that vendor, but the fingerprint doesn't matches, right. That also tells us that, hey, there's a potential issue here. I mean, there's a bunch of things that you can do just on a rules basis, like have we seen this email address before? Does the remittance information look like what we've seen before? For. But then, you know, you get AI that can do things like, well, what's the cycle or the timing of the invoices? Did you receive this invoice, you know, on a day or a time within the month? That makes sense. Um, is this the way you've typically been charged for, you know, this item? Uh, is this a weird item to see from that vendor? I mean, you can start to get like, you can go down this rabbit hole of doing more and more and more and more, and each one is incrementally better. But as you can see, like there's an inexhaustible list of ways to use AI to look for outliers that, you know, either fraud, clerical errors, overbilling, duplicate billing, you name it. Blake Oliver: [00:14:21] You collect all those data points. You feed them into the fine tuned model for that company, that vendor, and spot differences that a human would totally miss and prevent invoice fraud. Aaron Harris: [00:14:33] Yeah. I mean, you know, so so the reality is there's, there's, you know, fine tuned models around how an invoice is coded. Did. There's big general models that learn across sort of generally how things work. There's models that are fine tuned on a vendor by vendor basis that, you know, we one of the things that that I learned as we started this journey of building AI is that, like, you're not just building an AI that does a problem, you're building a system, and that system is going to have traditional tech. It's going to have AI. And usually when there's AI in it, there's a lot of different pieces of AI that work together to do that, to to end up with the outcome. Blake Oliver: [00:15:15] So those are two examples of AI currently in the product. The construction bidding example. Right. Helping us get the right price or figure out what the right price should be from a from a supplier. And then all this outlier detection around accounts payable. I think there was a third one that you mentioned. Is there another one that comes to mind. Aaron Harris: [00:15:39] Putting me on the spot. Blake Oliver: [00:15:40] It was AP construction. Aaron Harris: [00:15:44] Well, there's there's the supply chain. Blake Oliver: [00:15:45] Supply chain stuff. Yeah. Tell us about that. What is. Aaron Harris: [00:15:48] Yeah. I mean, it's kind of related to to accounts payable, but but the, um, what we're really looking for here is indications that, um, you know, in your supply chain, you're going to run into fulfillment issues, right? So, so something that happens, you know, out of the ordinary that triggers this, you've got some risk over here that maybe you're not paying attention to. Now, here's where I get to brag that we've invested enough in anomaly detection, and we're deploying it enough that I don't know the details of of everywhere. We're now deploying it. So, you know, you'll have to talk to Rob Sinfield, who's responsible for X3. But it's the purpose of this outlier detection is to pre-warn when you might have a fulfillment problem. Blake Oliver: [00:16:35] Okay. So customer is placing an order, and we're going to anticipate that we will not be able to deliver that order for a variety of reasons, right? Like ship shipment delays, weather, weather. Aaron Harris: [00:16:46] Right. And you're looking for irregularities in supply chain activity. Speaker4: [00:16:51] Got it. Blake Oliver: [00:16:51] Yeah. Well, I mean, that's really helpful because that's anticipating an issue that could cause a big problem with my customer. For me, if you can do that reliably. Right. That saves me a ton of like having to put out fires, I suppose. Right. That's what we want to avoid. Interesting. So okay, so those are some examples of AI in the product now, which is like part of a wave two. Right that you mentioned. And I think maybe it's worth sort of um, going back and just outlining these three waves of AI that, that you talked about in your presentation. So wave one task based AI, you mentioned that that's like. Aaron Harris: [00:17:30] Ai, that does, you know, a narrow, narrowly defined task in a very controlled way And oftentimes, you know, it's not even obvious there's AI, right? It's fully embedded in a workflow. Blake Oliver: [00:17:41] So categorizing a transaction right. That would be yep. You don't need Jen AI to do that. Speaker4: [00:17:46] No. Aaron Harris: [00:17:46] Um you can do it with Jen AI. But it's really expensive and it's unpredictable. Blake Oliver: [00:17:51] So then wave two that is generative AI. Um, and that's powering some of these these features that we discussed. And also Sage copilot. Speaker4: [00:18:01] Yep. Blake Oliver: [00:18:02] So we talked about sage copilot last year or the year before. Uh, and that was when it was just released. Right. It's been out for a year now. Speaker4: [00:18:10] Yeah. Aaron Harris: [00:18:10] So it's been a it's been a, like a cascading release. Um, we we first released it about a year ago with customers on our sage accounting product. We wanted to start with small businesses where sort of, you know, the realm of capabilities is pretty well defined and consistent and and a bit more narrow. So we launched with Sage accounting. We launched with sage for accountants. Right. So a lot of capabilities for enabling better collaboration between accountants and their clients. We've launched it for Sage 50, which we thought was a pretty important thing for us to do to demonstrate that, you know, we've got these kind of hyper modern cloud products, but there's also like a million customers on Sage 50. And, you know, they want AI to in fact, they're kind of demonstrating that they're hungrier for AI than we than we thought they were. Uh, I spent a few months now since we allowed our first customers to access copilot with Sage Intacct. And we've got two capabilities, well, three capabilities that were really focused on to start. Now, the first one probably won't surprise you. It's all around the clothes, right? So the idea is that you can interact with copilot who will constantly keep you informed on like what's what's what's preventing you from having your books closed, right? Maybe you've got some Pos that need accruals, or you've got something that's not reconciled, right? And then it helps you through the process of getting the books closed. One of the big promises of generative AI and Sage copilot is that because it can interact naturally and conversationally, we can now engage people outside the accounting team, right? So if you know so so think about me and my role. Aaron Harris: [00:20:03] And, you know, it's kind of a big company. Um, when I need to approve something, I've got to remember how to log into the ERP. I've got to remember how to navigate to to you know, where. The thing is, I have to prove I've got to remember what the how the screen works, where the button is. In other words, accounting products. Erp products are designed for professionals. They're not designed for the casual user. Um, so if you can sort of convert that experience to a conversation, then Now you've got something that actually engages me and can accelerate these workflows. So the second big thing we're doing is, is introducing these capabilities for engaging outside the finance team. And the area where we're starting is actually closely related to the close. We're starting with budget variants, right. So we're enabling budget owners to interactively understand how they're performing versus their their budget and to understand why, you know, there's these variances. So the third thing that we're doing with Sage Intacct with with Copilot is we're fundamentally training copilot with a deep understanding of how the product works so that, you know, if you're new or you're trying to do something you've not done before, rather than scan through help and maybe do like a Google like search, you just ask copilot, you know, how do I turn on multi-currency? Or how do I book a transaction in a foreign currency? How do you know? And it'll you have like a conversational, you know, with an expert right. It knows how intacct intact works. Blake Oliver: [00:21:31] Well, I have to ask you then, is it just going to tell me how to do it? And then I have to go do it in the interface? Or will it do it for me? Because that's something that I've, I've really want to see in all these chatbots that are being built. Speaker4: [00:21:41] When you say do. Aaron Harris: [00:21:42] It for me, do you mean like I gave you an example of turning something on? Speaker4: [00:21:46] Yeah. Like, say. Blake Oliver: [00:21:46] I need to turn on a setting. Speaker4: [00:21:48] Right? Yeah. Blake Oliver: [00:21:48] Uh, I need to change a setting in intact. Okay. I can use the Sage copilot to learn how to do it. But what would be better is if Sage copilot could just do it for me. Aaron Harris: [00:21:59] That's coming. I mean, so, uh, I had to step out of Dan's keynote today to to to do to an interview somewhere. Um, but what I saw in his his rehearsal, uh, what he was demonstrating was something they're working on where using natural language, you can configure the product. And the example that he showed was actually really cool. So within Sage Intacct we have these things called triggers. And triggers essentially say you know like identify something that happens in the product, like creating a customer, updating an invoice. And when that happens, there's some logic that you want to to, to execute, to do something automatically. And it requires kind of a technical understanding. Um, you got to kind of know how the product works, but you've also got to write some, some, you know, it's not it's not hard code. It's kind of like Excel formulas, right? Um, so what do you you know, what what he demonstrated was doing that interactively with copilot, where, you know, that trigger is created, it's written for you, and then you can go and test it. So we are we are working in that direction. David Leary: [00:23:05] So my brainstorm, because you see the Sage I thrown out everywhere. It's a little brand and logo. And then you have Sage copilot. It's its own brand and logo. And my brain is just like all your technology. And the copilot is just an interface to interact with the actual GL and the AI. It's really just a it's a UI, it's a UI layer. It's a chatbot. Blake Oliver: [00:23:27] Copilot is a chatbot, right? Sidebar. Speaker4: [00:23:29] It's more than that in the. Okay. Aaron Harris: [00:23:30] Yeah. So there's actually it's a really, really good question, right. So this is something that we've had to evolve with generative AI in the introduction of this conversational interface. You know, on the one hand, we needed to put a name on the thing that you now interact with. Right. And that name is Sage copilot. Right. And we're making a big deal about it. And you know, when you, you, when you go out there and you talk to customers and you generate marketing and you do your advertisements, it's got to be something that's kind of easy to understand. Copilot Sage copilot shows up in three ways, right? It's that conversational interface, but it's also some agentic stuff that's going on behind the scenes that's continuously looking for insights. The thing that customers love the most about copilot so far is when copilot prompts the user right with with with insights. And so there's this gigantic stuff that's sitting underneath copilot. Um, and then, you know. Speaker4: [00:24:22] You got to. Blake Oliver: [00:24:22] Be careful with that though. So it doesn't turn into Clippy. Speaker4: [00:24:26] Correct? Aaron Harris: [00:24:28] Yes. Well, so here's the good news. Um, our CEO used to be a grumpy CFO. Um, and he's gotten really involved in the development of Sage copilot. And the first thing that he said to the developers was, I don't want it to be cute. I don't need a friend. You know, I don't want it to ask me how my weekend was. I want it to be real, solve real problems. And most importantly, I want it to be correct. Right. It has to be correct. So if we ever if we ever show something that starts to make you think we're headed clippy's direction. Right. That's when my job starts to. To be in trouble. Blake Oliver: [00:25:04] Can you do it as, like an Easter egg, though, so that I could. I could enable Clippy if I miss. Speaker4: [00:25:08] What makes you think there isn't. Aaron Harris: [00:25:09] Already. Speaker4: [00:25:10] An Easter egg? Aaron Harris: [00:25:12] Do you know what Walid's background was? Right. Blake Oliver: [00:25:15] What was. Speaker4: [00:25:16] That? Aaron Harris: [00:25:16] Many, many years at Microsoft. Speaker4: [00:25:18] Right. Blake Oliver: [00:25:20] Maybe we could, you know, get some episodes of the accounting podcast into the product as a as an Easter egg. Speaker4: [00:25:25] So so. Aaron Harris: [00:25:26] I. Sorry, I really haven't answered the question right. So Sage copilot is the conversational interface, and it's what we choose to package around it into an offering. Sage AI is a much broader concept, right? It is everything that has, you know, AI in its ingredients, irrespective of how we manifest it. And so today I made this announcement about the AI trust label. Right. Where customers can, you know, when they're looking at an AI capability. Click a link and we'll pull up. Here's how we built it. Here's the models we're using. Here's what we're doing with your data. Here's why you can trust it. Et cetera. Et cetera. Et cetera. That is a Sage AI wide thing, right? Including copilot. So I think part of why this is important is that we've got products that are sold to sole proprietors, um, that just have to do making tax digital in the UK. And we've got large complex manufacturers running on X3 Sage I right. Is the overarching brand for everything we're doing with AI. Copilot is sort of how we package this up into this premium interactive experience. David Leary: [00:26:41] So you mentioned AI, and I feel like when everybody says that term, it's me asking the AI to do something on my behalf. But you kind of framed it in that concept of like, it's just doing stuff and it's going to surface to surface it to you through copilot. Speaker4: [00:26:54] I mean. Aaron Harris: [00:26:55] That's the real power, right? So it's the answer is it's both. Right. But but the real power of agentic AI is the fact that it can be autonomous. Right? That's the that's just kind of a scary word, right? Which means we've got to be super, super careful about the way we roll out this technology. And so, you know, the first, um, ebb and flow of this wave, if you will, right, is agents That only do work as prompted by a user. Right. Go off and find this information for me or something. But ultimately, you know, the real power comes in when it can be turned on with a job to do, and it's running in the background. Blake Oliver: [00:27:37] And that's wave three. Aaron Harris: [00:27:39] That's wave. Speaker4: [00:27:39] Three. Thank you for. Aaron Harris: [00:27:40] Getting us back on. Speaker4: [00:27:41] Track. Blake Oliver: [00:27:43] It's it's coming soon. I mean, you've already got some of these. Speaker4: [00:27:46] These waves overlap. They're overlapping. Yeah. Blake Oliver: [00:27:49] Um, so autonomous agents, fully autonomous systems. What is the timeline for that? And, like, what are examples, uh, of systems you help to build autonomous agents that you would like to see in the products? Speaker4: [00:28:07] Yeah. Aaron Harris: [00:28:08] So I'm going to make it hard for you to pin me down on a timeline. David Leary: [00:28:12] Um, we'll ask you again in six years. Speaker4: [00:28:14] Yeah, fine. Fine. Aaron Harris: [00:28:15] Yeah, you'll you'll actually remind me we did this podcast, you know, six years ago. And I said this in my in my keynote part to the partners yesterday, I spent a lot of time talking about the way we build AI and the way we deploy it and the way we sort of roll out capabilities. But I started that conversation by putting up an Axios article that was the results of a survey with businesses where something like 75 or 80% of businesses, quote, want AI companies to take it slow and get it right. And that's why I'm it's going to be hard to pin me down on these things. Like I can't tell you when we're going like what date we're you know, the Dan is on a deadline to now turn on, you know, Sage copilot for another 10,000 customers. It's going to be when we've got the confidence that the experience is right, that it's that it's accurate. Like, I know I'm giving you a super corporate answer. Blake Oliver: [00:29:10] No, it's totally. Speaker4: [00:29:11] Cool, but it's true. Blake Oliver: [00:29:12] Fair enough, fair enough. I mean, you know, it's hard to we develop an app, and I know that it's hard to even forecast beyond 2 or 3 months. So let's just talk about like what you envision. Like what would you like these autonomous agents to do? Like what? How could they help me as a CFO? Aaron Harris: [00:29:29] Yeah. So so so at the highest level, right. Well, we've been essentially telling the industry is that AI is going to automate all the low value, but nevertheless non-negotiable work that CFOs have to do. Um, on the one hand, and then they're going to enable them to be more successful on the more strategic stuff that they're going to use with their free time and their free resources. So on that non-negotiable stuff, you know, everything related to the repetitive accounting assurance, um, and reporting tasks, I expect to be able to be done by agents. Um. Now, the critical thing is getting the experience right. Right. Blake Oliver: [00:30:11] So what are some examples of those. Aaron Harris: [00:30:13] Uh, end to end accounts payable? Right. Uh, receiving a bill, entering the bill. Coding the bill, overseeing the approvals required for the bill, creating the payment request. And, you know, given sufficient sort of, um, trust by by a supervisor paying the paying the bill end to end, right, where a human is involved, you know, only when there's something where the AI doesn't have enough confidence or it's sort of significant enough that you want your supervisor to take a look at it. Right? Don't send out $1 million payment before, like, I take a look at it. And actually, the other thing that I think is really, really interesting is, you know, we've learned to accept in the accounting industry these like super complex but nevertheless very rigid workflows. Right. Like, what is the workflow to approve an invoice when the amount is this when it's on this project it involves these departments. Blake Oliver: [00:31:13] Oh yeah. We've got a whole. Aaron Harris: [00:31:13] Flowchart above this and this line item. Right. You know, so. David Leary: [00:31:17] There's entire apps that only do that. Yes, only does that. Aaron Harris: [00:31:20] Yeah, yeah. I mean, kind of archaic when you think about it, right? So, you know, one of the things that I'm hopeful for is that I can now start to make this workflow quite a bit more manageable, um, more flexible, where I is making some risk decisions. Now, it's not making these decisions without some guidance. Right. So, you know, hey, here's the things that I care about, like pay invoices from these vendors on time because they're critical to my supply chain. Like, I can probably figure that out. Um, so, you know, there is going to be some guidance to it, but I really do think the day is coming where these really rigid, complex workflows are replaced by something smarter. Right. That figures out, hey, who should look at this? Who should sign off on this? And frankly, I think it'll it'll actually result in a better outcome. Um, you know, I probably shouldn't say this in front of, I don't know, like your millions of listeners, but by the time something gets to me to be approved, it's gone through six approvals. Blake Oliver: [00:32:26] That's a lot of approvals. Aaron Harris: [00:32:27] Yeah. And so when I approve it, there's two things that I'm wanting to understand. Number one, I'm not going to I'm going to I'm going to actually message my finance partner and I'm going to say, hey, have you looked at this and are you okay with it. Right. And then actually I'm probably like, if it looks like something that I know is happening, I'm going to approve it. Whereas imagine now that occasionally I get prompted by I, hey, there's something going on, uh, in purchasing within your organization that I think you should look at, and here's why I think you should look at it. Like, that's going to lead to a much more engaged experience. Um, you know, and I'm going to take a much more critical look at it. Blake Oliver: [00:33:07] I was just at another conference, uh, focused on, uh, travel and expense management and, uh, talking about. We were talking about all the different possibilities for AI in in that area, that one area. And a great example was looking through all of the purchase orders, all of the purchasing for a large company like Sage and just focusing on software subscriptions, because how many times are there duplicate subscriptions? Aaron Harris: [00:33:36] Yeah, it's a huge it's a. Blake Oliver: [00:33:37] Plague. Aaron Harris: [00:33:37] Right? Blake Oliver: [00:33:38] I mean, how many seats go unused for all these apps? Aaron Harris: [00:33:41] Yep. Blake Oliver: [00:33:42] That's something that, like a human could do, but it would be somebody's full time job, you know, just to go through those subscriptions and try to consolidate them and. Aaron Harris: [00:33:51] Yes, but but but to do this, you really properly need an agent that's got the ability to work in a really messy system. Yes. It's got to traverse, you know, all parts of the organization and then sort of understanding, well, how within this part of the business do I gain an understanding of what tools they're using. You know, you can think about one of these like workflow process mapping products like on steroids. So huge value in doing this, but there's no way to do it without agentic AI. Blake Oliver: [00:34:22] So you painted a picture of an autonomous accounts payable agent. What about when it comes to the financial close, the month end close? Aaron Harris: [00:34:31] Well, we're getting rid of that. Blake Oliver: [00:34:32] Yeah. So so let's talk about that. You mentioned this trust label for AI. What is going to allow us to trust what the AI is doing so that we don't have to do the traditional financial close? Aaron Harris: [00:34:46] Yeah. I mean, so, you know, getting serious this is the big challenge of AI, right? You know, getting the psychology right. You know, of the relationship between the user and the technology, um, you know, understanding what what's the trust journey that the user has got to go through to increasingly hand off responsibility to something where that user is still accountable for the outcome. Right. You have to understand that that psychology in order to design this experience. And so, you know, I think it's I think it's a journey, uh, you know, I think, you know, the way you design the experience is one where you increasingly enable more autonomy on, on the agent's part. And I think the way the trust will be earned to enable that, you know, is through a lot of transparency from the AI on what it's doing, how it's doing it. You know, why it did it, right. So, you know, like, am I going to trust AI to handle accruals for me? Well, probably not until I see some of its suggestions for how to accrue things. Not until it proves to me that it actually reached out and spoke to some purchaser. Um, not until like, I can be confident that it understands, And, um, you know, when invoices come in and, you know, are there things that haven't been invoiced yet potentially that should be accrued? Like, I want to I'm going to want to see. Blake Oliver: [00:36:19] Yes. Aaron Harris: [00:36:20] Right. In a very transparent, auditable way what the AI is doing before I say yep. Okay. Like you can do it now. David Leary: [00:36:27] So when it comes to this trust. Do you guys have like, this little badge or whatever on some feature or answer that's given. And you have this trust label. The trust label. Now is this something that you guys are awarding some sort of score and blessing it, or is this like feedback from, you know, Netflix in the old days? Every time I get a DVD delivered, you'd say yes, it came or no, it was broken. And that's how they can measure the quality of all the deliveries. It's something you're aggregating across all the users, like a bunch of people are using this bill scan for. These are the, what do you call it, the the laser focused. Aaron Harris: [00:36:59] So invoice process. David Leary: [00:37:01] Yeah. The invoice processing. Aaron Harris: [00:37:02] Yeah. So this isn't necessarily intended to provide a score for trust. It's it's really meant to provide A transparency. David Leary: [00:37:10] Or the audit trail. Aaron Harris: [00:37:11] Right. So we're not saying like here's how much you should trust this. We're saying here's here's the compliance that we are subject to and we are meeting. Here's the framework for security that that we're following to ensure your data is secure and private. But also here's the models we're using. Right. And are we using your data to train models. Uh, what are so what's the process that we follow to ensure that, uh, you know, we're we're building this safely and ethically, and some of those will be consistent across capabilities. But a lot of it like, you know, what are the models we're using? Um, what are the implications for data security and data privacy are, you know, are we involved in training models and and how do we do that? It's going to vary on a, on a feature by feature basis. And so, you know what I, what I, what I've seen talking to customers Is that there's varying levels of sensitivity to AI. You know, there are some extremely cautious organizations, especially in financial services, um, where, you know, the out of an overabundance of caution they won't use, for example, invoice processing, because one of the models we train on the collective right users. Um, so it's perfectly safe, right? And I can explain why it's safe, but but we're enabling them to make the decision, an informed decision about whether really. David Leary: [00:38:41] Disclosing what's going on behind the curtain. Right. I mean. Aaron Harris: [00:38:44] We're kind of comparing it to a, a nutrition label. It's kind of a way to think about it. David Leary: [00:38:48] Now, do you see a march of that going beyond Sage products to all the third party ecosystem, other apps that integrate with with you that are using AI as well? Aaron Harris: [00:38:57] So we're encouraging other AI creators to, to to align with us and collaborate with us and come up with an industry standard The if, if, if the regulatory environment had progressed far enough. Actually, what I would love to do is just throw some sort of kitemark on my eye that says, yeah, this has been this, this has been certified to meet government regulations and standards and, you know, by this third party auditor that certifies it. I think someday we're actually going to get there, but it's not coming fast enough. And I've got the problem today is I need our customers, prospective prospective customers, to be able to make informed decisions about whether they want to use the AI in the product. Blake Oliver: [00:39:44] So Steve Hare, uh, in his day one keynote, mentioned that the auditability of all the AI is super important to you. Aaron Harris: [00:39:54] Yeah. Blake Oliver: [00:39:54] Um, and he's a trained accountant, so he wants to know, like you've been saying, what is the AI thinking? How did it come to its decision? It needs to show its work. That's how I take it. Yep. So let's use the example of an AI agent that's doing accruals. I here's here's an invoice that came in I paid this invoice and now I want to accrue it or defer it or whatever. What have you. And it's got a you know, make the work paper do the journal entries. So you're going to I take it you're going to show me then it's thought process for all of those journals. I'll be able to trace back the thinking. Aaron Harris: [00:40:31] Yeah. So, uh, you know, if, if, um, if you in using sage Copilot, you ask a question, you know, like, what are my top ten customers by revenue? There's there's a, you know, explain how I arrived at this, and you expand that and it'll take you through, like, here's the process. Here's the steps that I took to arrive at this answer. I think the what's critically important there is also, you know, here's how you can verify. Blake Oliver: [00:40:56] Yes, that was always the issue when these, um, chatbots first came out, was it would give you an answer. Yeah. But there was no way to verify that. Aaron Harris: [00:41:04] But I mean, you've watched the evolution, right? It's it's it's so rewarding now that when you use one of these research capabilities, they're actually like footnoting everything, right? You know, they're sourcing everything. So you can and actually in that way it's not just creating more confidence on on my part as a user, but I actually find that I want to go dig more. Right. And I'm going to those sources. And I think that's the case with CFOs who are interacting with Sage copilot. You know, they don't just want an answer, right? They want the ability from that answer to go and do their own research. Right. They want to dig deeper. They want to include other people so that verify, you know, thing. Yes. And there will be, you know, some other iterations and ways that's manifested. Um, the idea, though, is that you've got to enable that customer to go to the next step in the process. David Leary: [00:41:58] I texted Blake about halfway through your keynote, and I was like, tell us your you built your own, uh, language model without telling us you built your own language model, and you kind of made an argument of how the existing off the shelf models are kind of bad, and you don't want. Aaron Harris: [00:42:13] To you maybe. David Leary: [00:42:13] Don't want to use them versus using an accounting specific one. Blake Oliver: [00:42:17] They're very expensive to use. Aaron Harris: [00:42:18] Well, yeah. So so two things. Um, they are very expensive. Um, but but the bigger thing is they're very good, but they're not outstanding. They're very good at, at, at, like, the widest possible spectrum of things. Uh, I don't want it to be very good at the widest possible spectrum of things. I want it to be an expert at a very narrow set of things. I mean, relatively right. You know, the industry of accounting and compliance. Right. Um, and so what's happened over the last couple of years is the emergence of these much smaller models that can be fine tuned. And so they've got all kind of the conversational capability that the very large models have. Um, but they don't have that same corpus of knowledge that the large models have. And so you've got the ability to we they're called Laura's. Right. You're able to add these layers on top of these models. You know, a large language model is already a, you know, a, a neural network of, of layers of nodes that are connected. Right? We just add our own layers on top of that, that, that have this specific knowledge that we want to to, to really be the focus of the way it works. That wasn't possible right until until a year or so ago. David Leary: [00:43:39] And then you're working with the AICPA? Yes. Build the accounting model. Aaron Harris: [00:43:44] Yeah. Blake Oliver: [00:43:45] So that one will then know how to do the accruals and deferrals according to GAAP. Yeah. Is that the idea? Aaron Harris: [00:43:50] That's that's the idea. Yeah. I mean we can get that from accounting textbooks. Yeah. Um, but but I mean, there's so much more that you get from the AICPA. I mean, fundamentally like, you want this without giving it the same trust you would give a CPA. Like you want it to be as capable as a CPA. And the only way to do that is to train it on the material that's used to, to train and certify CPAs. Blake Oliver: [00:44:17] So one of your competitors. Aaron Harris: [00:44:21] Unnamed, I guess. Blake Oliver: [00:44:23] Well, let's say it, let's name them Intuit QuickBooks has released four AI agents. Oh yeah, this month they're coming. Aaron Harris: [00:44:29] Yeah. Blake Oliver: [00:44:29] Um, in June, as we record this, uh, agents that will do accounting and finance, and I forget the other two, David, but they're they're going to be somewhat autonomous, it sounds like. Um, how do you contrast, like, your approach to, like, the approach of, say, a QuickBooks or Intuit is. Aaron Harris: [00:44:50] A lot less reckless. You know, when they launched their first agents, they made all kinds of headlines in the Washington Post and Wall Street Journal. I mean, they're there. Agents were there. There I was trained on I was actually really shocked at how they trained these agents. It was trained on community content. So you're sort of relying on the community? Blake Oliver: [00:45:15] Yeah. The forums. Aaron Harris: [00:45:16] Right. I mean, and they were just they were just doing rag, right? They were, you know, sorry. Um, they were they were the I was querying the content as needed as opposed to training the content into into the model. Mhm. Um, like, I'm not going to say that Intuit can afford to be more reckless than we can. But what I can say is, you know, we're we're one of the only companies I know that is like, ruthlessly focused on the accounting profession. Like, we sell Sage Intacct to a CFO, like full stop, right? There might be some other title that that, you know, subs for CFO, but that's who we're selling to. And we know that that CFO needs to be trusted, right? And they're not going to use something that they don't trust. So we've made loads of progress on building Agentic capabilities. A lot of the way Sage copilot works today is, in fact, Agentic. But we're not going to release fully autonomous agents until we are absolutely confident that we're not going to erode the trust of our CFOs. Blake Oliver: [00:46:25] We're going to take your time, do it. Aaron Harris: [00:46:26] Right, get. Blake Oliver: [00:46:27] It right and get it right. Aaron Harris: [00:46:28] Yep. Blake Oliver: [00:46:29] Well, I think that's a great way to wrap things up. Aaron Harris, thank you so much for joining us on the podcast. Aaron Harris: [00:46:36] Thanks for having me again. We'll we'll follow up in a few years and see how things worked out. Blake Oliver: [00:46:41] Hopefully sooner than then. Aaron Harris: [00:46:42] Sounds good. David Leary: [00:46:43] 2031. Aaron Harris: [00:46:44] Super 2026.