Podcast

Today we're talking about AI agents, supposedly the most disruptive AI technology we're gonna see in the near term. Shared my thoughts about them in the past, but we're gonna do a little check in a few months into 2025. Where are agents at when it comes to accounting tech? Not like a ton of like earth shattering applications yet, but we actually know a whole lot more now about what those applications will actually look like. And last week developers got a huge new set of tools to build cool agentic stuff into the software we already use. So today we're rounding all that stuff up. By the time you, by the time you head out the door, you should have a pretty good picture of where we start seeing more AI agents and accounting tech. Let's talk about it. I try not to get super hyperbolic when I talk about AI things. It's always worth acknowledging. It's a huge spectrum of where people think AI will ultimately take us from. Pretty much replacing all forms of knowledge work. Like that's the ceiling. So it's not even accounting specific, it's the ceiling for what AI could do. May just be replacing all knowledge work. The floor is kind of what we have now, and increasingly the floor, frankly, it's really juicy because we're, we're a long, long ways from the tech that we've already got being implemented into all the tools we use in a meaningful way. Like there's a good five years, a really meaningful software development to be done on top of all the language models and all the things that we've already got. But far and away what people say is, is likely the most impactful thing AI will do is, is more sort of agentic workflows. So just to kind of define that vernacular, an AI agent is something that's going out and it's doing stuff. Like if you think about an AI assistant being something you chat with, you interact with, oftentimes you're talking about doing the work. Whereas the AI agent going out to a computer, it's doing the tippy tap and it's clicking around on stuff the same way a human user would. It's both really cool and really exciting, but also kind of dumb in other ways. So it's cool from the standpoint that then an AI agent could do anything theoretically that a human user could do because it uses all the same interfaces as a human. It's kind of dumb on the other hand because it's like, man, isn't that a wildly inefficient way of a computer trying to do Something like, like navigating this, this thing that is designed for like the visible range of light that humans can see and for our very limited like input peripherals, the keyboard and the mouse. Like really, that's the, that's the best way for a computer to read a thing. I've long thought this about like paper, physical documents. If you think about a bill, like a printed invoice that is made in a way where a human can look at it and read it and understand it. And if you think about the, the amount of investment and work and thought that's gone into. How do we most accurately ext extract all the human readable things from this document? Man, couldn't we have just reduced that to a barcode a really, really long time ago? Something that the human will not bungle. Sorry. That the computer will not bungle when it's trying to extract all the details. Think about, I don't know, books, I don't know, things you buy from the grocery store. Imagine if on the back of every item in the grocery, grocery store was like a paper invoice and you scanned it and something did some OCR technology and it's like, I think it costs this much. We didn't do that. We made a barcode standard and you run the little thing and it goes boop. And you know exactly what it should cost. Why we haven't done that for receipts, for invoices. No, I have no idea. We wouldn't even need OCR if we had like an agreed upon structured data file that could be extracted from a QR code or barcode or something like that. And AI agents, they kind of feel like the same thing. Like we're forcing them to navigate these very optimized for human interfaces. And the result right now in the current state of things is they're, they're still very slow, pretty dumb. But now, and this was not the case a few, few months ago, we have like a AI agent, a, a generalist AI agent on the market that anybody can go sit up and play with. It's available on the $200 a month ChatGPT plan chat. It's called ChatGPT Operator. If you haven't seen it. Operator basically has a built in Internet browser where I say go and do this thing and it uses its built in Internet browser to go and do that thing. Interestingly, I can cut loose up to three operators at a time right now to go and do my bidding online. Pretty cool, like amazingly novel. Like you can point it at a Google sheet or an Excel workbook that is on Excel online, and you can, like, see it working along alongside of you. Like, it's pretty wild. Kind of opens your eyes to, for one thing, multiplayer software experiences. So online collaborative document editors being an example, like, like spreadsheets or Go Google Docs, online whiteboarding tools like Miro or figjam, any online software where you can see what the other users are doing at the same time. Doing that alongside an AI agent. Super interesting. Like the feeling of proximity. And so if you just look at something like using Excel, like, we're not that far from a reality where if I need help with a formula or something like that, like an AI agent can function as a tutor that actually goes through those things with you and you kind of see its hands working. In fact, the version of this that just makes the most sense, where Microsoft has again crapped the bed on all things AI, is the AI agent that's built into Windows, that's built into the operating system. I think we're real close to a world where we are vocally talking with AI assistants who can see and understand the things that are happening on our screen. Maybe the closest thing we have to that right now is like, you can send a screenshot of the desktop to the ChatGPT desktop app and talk about it, but I'm talking like, streaming the video from your screen so that at any point in time, you can have a conversation with an AI assistant about it. Google, interestingly, has some tech that will let you do this, but you kind of have to bodge it together. And it's not great. Like, it can see everything in the browser as you're, as you're talking with it and have conversations about what you're working on. But we are, we are very close to a world where throughout the day, I'm probably having a running conversation with the AI while I'm working on something, and they can see what I'm working on. And so just like having that poor little junior sitting in a chair next to you as you work, and you're like, go get that email off the printer for me, or whatever it is that you ask them to do, we're going to have that AI sitting there that can run off and research a thing for you or troubleshoot why a thing in a program isn't working. And that's going to be really cool. And like, part of the really exciting promise of agents is that theoretically you can cut any number one of any number of these things loose at a time. So ChatGPT operator right now you can cut three operators loose at a time, but there's nothing to stop you from cutting loose a million operators besides cost and resources and all that. So what you and I are competing with down the road is not like this one kind of one dimensional conversation we're having with ChatGPT, for example, if I'm researching something, what I'm competing with is actually in a like entire team of AI agents going out and researching something simultaneously in like a collaborative way. Like that's the future of AI agents and that's super cool. But the best version, like the best thing we have on the market that normies can use, that doesn't require developing something custom yourself, that's ChatGPT operator. And it is really interesting to like kick the tires on it to try it out, but it's very slow, it gets confused easily. And so right now it's not in a state where it's like going to be very helpful to your business. I will say one very smart thing that they did is that when operator finishes a workflow, you can say, save this workflow and then you can run it again in the future. And it's like it's going to be smarter in its understanding of what steps it tried previously that didn't work before. Ultimately what was the solution that did work. So as you watch it work, you'll find it. It'll take a real meandering path trying to figure different things out. But if you save the workflow theoretically, then the second time it does the same thing, it has a better understanding of how to get there directly. And I mean, even if the version one is really, really dumb and really, really slow, not unlike it is with a human being who does something for the very first time, the promise of these operators or AI agents is man. Even if it was just a little bit better and I could refine that workflow so that it can do it in a reasonable amount of time. If that's a repetitive process, once we've got that set up, we can do that until the end of time. Like in theory, a human never has to touch that again. Right? Like that's a big deal. So it doesn't have to be as fast as a human or as perfect in many ways. It just needs to be good enough. And we're getting close to what good enough could be. This episode is sponsored in part by the folks at Canopy Canopy, the Practice Management System. I've spent a lot of time digging through PMs trying to filter through the 43PMs. Now that I believe are marketing to North American firms. As you probably know, Canopy is on my short list of recommended practice management systems. And nobody likes having any more tools than you absolutely need to have, right? That's why you start building your tech stack with the practice management system to let that do as many things as possible. And when it comes to all in one Erskine pretty hard to beat Canopy. In fact, a couple years ago we did this giant roundup of feature comparison of all the PMs on the market, Canopy may have actually had the most complete feature set of any PM out there. Now that's not always the answer, right? Just more, more, more. But I can tell you we're also pretty burned out of just like having to build these ridiculously big stacks. So having a having a PM behind your firm who has the mindset of we want to try to pull everything that we can into one platform for you and to know they're going to keep investing in that in the future, I get behind that. I'm hey, I'm a big time tech nerd, but I also don't like playing software night janitor. So if I can get all that stuff in one place as much as possible, I want to do that. That's Canopy. To learn more about Canopy, check out the link down in the show notes. This episode is sponsored in part by Team up, who helps you directly hire talented accountants in the Philippines. Why do you hear so much about the Philippines specifically? Used to be like India, right? Why is the Philippines caught on A few things that I'm stealing from this excellent salary guy Team up put together. You can swipe that down in the show notes. Four reasons for why the Philippines has caught on 1. They've got a critical mass of talent, loads of people there with decades of experience and also just coming out of school. It was the only time in my firm running journey where I knew I could start the hiring process and have somebody within 30 days. In fact, I just talked with someone who did that like within 14 days. So the talent is there. 2. English fluency Very good English speakers will have an accent usually, but very, very common to have strong English speakers in the Philippines. Cultural compatibility Very strong Western influences in the Philippines. So generally easy from a work culture standpoint to get connected with and then a work ethic. So Filipinos pride themselves on bringing energy and optimism to work, which isn't always the case when you're hiring somebody, right? So to have that enthusiasm on your team to provide an opportunity to these folks, an awesome career, a better path, opportunities for their families. Like they turn up to work with just an entirely different level of excitement. And that was my experience. Anyways, Team Ups put together a great guide for how to get started with this stuff. Swipe that one down in the show notes. Now second update, a tool called Manus. We've talked about it a little bit, it made some headlines last week. It's just, it's a very, very good research agent. It'll go out and do kind of a deeper version of what ChatGPT's deep research does, which you know, I'm a really big fan of. And basically it works by cutting a whole bunch of agents loose to research something in like a collaborative way. It has a bunch of built in tools for doing things like writing and executing code if that's necessary to solve the tasks that you're asking it to do. But that was like, in many ways that was like Operator 2.0. It was like, here is kind of what the next more powerful version of Agen can look like. That Manus agent, it can run for basically up to 90 minutes at a time working on a hard task. And I think what it will ultimately do is probably just put pressure on open AI to ship a better version of Operator. So Manus, it's from a Chinese company, not something you're ever going to be plugging into your organization in like a west, the western business setting. But it is, it is a good peek into what's around the corner. And then the third thing I wanted to bring up, which I'll admit is a little bit nerdy, you may have seen talk of this online MCPS or Model Context Protocol. So MCP is a framework introduced by Anthropic. Anthropic is the company behind Claude. If you've ever used the Cloud AI Assistant, it's very, very good. And the reason I'm bringing up MCPS is on the back end, on the developer side. It represents a pretty big unlock or opportunity to make AIs work with our tools more effectively. So what an MCP is is something that the software company creates that basically serves as an interface between their program and the data that lives in say your practice management system or accounting ledger and a large language model. So right now we've got really smart LLMs, but they're not connected to all of our systems. Right? And that's a big pain. And to have to go out and build out all these manual connections to every single piece of software that there is when every single one of them has a different API, like a different way of accessing the data in the program. Oftentimes the APIs are incomplete where, you know, for example, very commonly with practice management systems, you can programmatically get access to, say, the context contacts in the system, but not the projects and all the other data, the reporting and stuff like that. The idea behind mcps is it's like this vastly simplified interface where a language model doesn't have to know the exact language and vernacular of how to make an API call for a bill in QuickBooks. Rather, the MCP handles all of that. And so if I'm having a conversation with a large language model, it can then speak to the MCP for, say, QuickBooks. And if I ask it for, you know, my largest overdue bill, it knows how to make sense of that request for like, the data structure that is behind the scenes. It kind of hides all that away so that it's a LLM friendly interface. And this means if I wanted to build an app that pulled in data from my practice management system and my ledger and a couple other, you know, accounting professional tools, then rather than having to, with my software, build out support for every single one of those APIs individually, I can instead just build out support for their mcps if those software companies choose to develop them and make them available. Which then vastly simplifies the notion of my tool being able to do cool stuff with, with AI and, and in kind of a conversational way, get data back from those systems. Now it's also a pretty big competitive decision for, say, our practice management systems to be like, are we actually going to stand up an MCP and let other apps access it? You know, right now, like, we do not have native integrations across tons of the important tools that we use, like look at tax intake tools. Like, that's become such a powerful, juicy part of workflow that practice management systems aren't building out direct integrations. It's too early to call whether that's like enough of the secret sauce to where you need to do that yourself or whether it's worth, like, building out integrations like partnerships there. And so it's by no means a foregone conclusion that everybody's going to run out and build these MCPs that will make it easier for language models to work with the app. But particularly as agents get better and we increasingly rely upon agents to do some of our work. Like, there's a world in the near future where, you know, a member of your, of your organization and your practice management system may just be an AI agent and you assign tasks to that AI agent. And if these agents actually do what it kind of looks like they're going to be able to do, then increasingly it will be the agents using many of the software platforms we use in addition to us, and so given to practice management systems, for example, that are otherwise equal. If one has an MCP and the other doesn't, and that means that, you know, the AI agent that I'm working with, maybe it's ChatGPT operator, maybe it's something more accounting specific. If that agent works with 1pm but not with the other PM, that's going to make it pretty hard to not go with the PM that is supported or that has an MCP so that I can do more automated stuff on it, Right? Similarly, for the accounting ledger, and this is maybe even a bigger, more interesting discussion, especially as Digits now has launched kind of their QuickBooks competitor is will the ledger companies stand up MCPs? Historically, the accounting ledgers have been pretty ecosystem forward where like they want to build out an ecosystem of apps that connect to the ledger because that kind of gives them a competitive advantage, but they only want to do that like to a degree. Because if then people start kind of sidestepping maybe some of the premium features of a QuickBooks, for example, by building something on top of it that can interact with it via MCP, then is that sort of like stealing QuickBooks Thunder of what, what they could potentially ship from an AI standpoint? It'll be super interesting to see. It is like it kind of kicks off a whole new complication and how you decide to either support the ecosystem or lock yourself down and not interact like not integrate with anything else. Because you can see a world where there's a collection of tools that are all like MCP AI first that all work together really nicely. Let's say I got a practice management system that AI agents can work great with and a ledger platform that has an MCP that I can work great with. If my whole stack does that stuff really, really well, that unlocks a ton of workflow opportunities. And all it takes is then one tool in my stack of tools that doesn't play nicely with AI stuff to totally gum up my workflows. It is really the difference. Like if you're thinking about an AI agent using applications, it's the difference between ChatGPT operator having to navigate through an application manually like a human user does and kind of learn it for the first time, versus that same agent just being able to ask the MCP for the same tool. Hey, can you get me the largest outstanding bill or whatever it is it needs and immediately getting that back. So the more tools that are reliant upon having to navigate the user interface, the less powerful your AI agent is going to be. Because that's the fallback scenario. Having to navigate a program the way a, you know, out of style, slow, boring human would have to by clicking through the user interface. So it's going to be a fascinating next five years in accounting tech. There's going to be folks that come out of nowhere that are really suited for the AI stuff and especially are suited for working with AI agents to automate workforce. There's going to be incumbents right now who are huge, who will see this like the changing tide too late and people will bail on them because they will see more agentic, more AI friendly alternatives. Looking at you ledger players could very well be what gets folks off of QuickBooks and or Xero. This MCP framework came out last week. Another big thing that came out last week is OpenAI basically created a system that developers can use that is basically letting them run with their own version of Chat GPT Operator, their own version of building agents that navigate the web. And this is, this is a big deal because it's like all of six lines of code for you as a developer to now build your own version of this. And so if we look at ChatGPT operator, that ought to be the very worst AI agent we ever get and the worst version of it that's out there because it's a generalist agent, it's not given any specific skills for specific websites or anything like that. Whereas if I'm a developer, I can now go build a version of that that's optimized for QuickBooks or carbon or whatever the kind of standard set of tools that small accounting firms use are. I can build a version of that with a little bit more context that's optimized for navigating the common apps that we use better than a generalist agent like ChatGPT operator. That tool set for developers just came out last week. So get ready to see agent stuff getting bolted into all sorts of different apps. Get ready to start hearing about mcps. And some software companies like touting the fact that they're leaning into AI in a way that many other companies and slower movers are not. Interestingly, you know, we've always had technological progress and software is getting better over time. I think what could happen over the next few years is the rate at which the leading platforms are outperforming kind of the laggards that Delta is getting bigger. So if we really get to the point where like, AI is a big lift here and AI agents start to become a little bit more capable, the output, like the productive output that we're capable of with these tools versus without these tools, it's probably bigger than that Delta has ever been before, which makes the software decision bigger than ever before between platforms that are leaning into this stuff and platforms that are not gang. If you're, I mean, if you're a tech nerd, it is a spicy time to be an accountant. A lot of people still like, concerned and confused and not knowing how to feel about kind of the future of work and how AI impacts or doesn't impact accounting. We're going to talk a little bit more about that next week. Like what stuff genuinely is on the chopping block because some stuff is and sort of the, the, the trends that we should expect over the next 12 to 24 months. This was a nerdy one, but just wanted to do a quick check in on the state of AI agents because it's kind of the biggest, it's kind of the holy grail of AI applications is if they can get agents to work well, then they're going to like be running around in the apps just like we are and that will definitely change the way that we work. So nothing that is today like a have to have, like you're not missing out on anything by not going out and burning 200 bucks a month on, you know, chat, GPT, the pro plan. There's nothing like that yet. But from an infrastructure standpoint, we've gotten a couple of really big developments that have made this easier than ever before for the folks inside the accounting space to start plugging this in and having a bit of a play with it to see where it could take us. Exciting times. You better believe we're going to be here talking about it when it does happen. Got any questions? Things that I explained poorly almost certainly Drop that stuff down in the comments. We could fill it out further in the next Q and A. I'll see you there. Sam.