033 Another Week - Another Great AI Announcement for Accountants

Okay, today's the day we're talking about the AI release. I've been teasing since last week that a practice management company in the accounting space made. That is the biggest, the biggest like most exciting snapshot of the future that is to come for accounting tech. Let's talk about it. It's cool. Today on Jason D. Okay. Company called Pixie. They are a UK based practice management system. Just cut straight to the chase. They announced a quote unquote AI agent. And so today we're going to talk about agents which are fundamentally different than the AI applications that we've seen in accounting software thus far. Ultimately they are the most exciting long term application of AI for us. I didn't think we would see one this early, but there are some strings attached to this launch from Pixie that we'll talk about. But to put all this into context, let's talk about agents and how they are different than like ChatGPT, you know the generative stuff like we've seen thus far from Cat, from, from Carbon and from Canopy. To date most of the AI applications we've seen have been like just stuff built on GPT. So for example, Wendy's just announced an AI chatbot that where like customers can talk with its chatbot and it'll take their orders, do like a whole back and forth thing and 80% of people who tested it preferred it to having to talk to a human to take their order which is interesting. There's also like a bunch of different spins you could do on that Wendy's chatbot where maybe like the, the customer doesn't just talk with that Wendy's chatbot but you have a chatbot listening to catch any errors. Like that's a higher touch version of it. We've talked a lot on this channel about how accessible building your own chatbot has become. GPT4 unlocked that in a really big, really exciting way. I mentioned the other day. I think we'll see more influencers creating their own chat bots to enable people to connect with that person in a really scalable way. Got an example of this? Just saw this news story this morning. Snapchat influencer launches Karen virtual girlfriend bot. I'm not even going to link this in the show notes because it doesn't deserve your click. 23 year old Snapchat influencer used OpenAI's technology to create an AI version of herself that will be your girlfriend for $1 per minute. The future is here my friends. I mean the good and the bad part of this, that's like exactly what I expect to see more and more of. Girlfriend thing, probably, but more like thought leader, like a chat bot to give you access to that thought leader. And like their stance based on kind of all the context that they've developed through their content and all of that. It sounds kind of weird right now, the notion of paying for a chatbot to get access to somebody's expertise. But it's coming and it's just gonna keep getting, I think, more normalized. But right now it's still pretty weird, especially if it's to be that person's girlfriend. So those are all examples of generative AI GPT doing genuinely helpful things. But agents are a big step up from that. So some examples of agents that you may have seen on Twitter or wherever you hang out online. AutoGPT, a new project Microsoft's working on called Jarvis. Those are probably the two biggest examples of agents that have gained popularity in the last few weeks. And this stuff is like maybe a month old now, so it's still very early days. But it is exciting to think about what this stuff is capable of. So the difference between an agent and just the generative stuff we've been working on is an agent is assigned a job and it will work on that job autonomously until the end of time, until it can complete it. And so to go through one example of a AI agent framework, the most popular one right now is called AutoGPT. It's open source. Anyone can go out, download the code from GitHub and run it. Now there's a bunch of browser based systems where you can run auto GPT. So kind of like chat G, you can run AutoGPT just from a browser and give it instructions to do stuff. It's pretty janky and not like super useful yet, but like, you'll see the whole framework of how agents work, like, oh, this is actually what the future of work looks like. So the way Auto GPT works is you give it a set of goals and it will create its own subtasks to achieve those goals in like this infinite automatic loop until it completes the objective. And we've seen, we've talked about some versions of this in past episodes. So if you remember, a while back we talked about a study that was released by OpenAI's Arc. It's like their accountability group to ensure that this stuff doesn't go full evil. And basically they gave an AI bot full access to the Internet and they gave it a bank account. And this is not like what influencers are doing on TikTok. Right now, where they're like, I follow ChatGPT's advice for 48 hours. It's not that this thing genuinely had money and it could go anywhere on the Internet. They cut it loose and they said, okay, we just want to see what it will do. And so they monitored it and basically said make a bunch of money. Gave it 40 bucks. And one of the first things it did was went out to try to invest that money. It ran into a captcha, the little thing that checks to see if you're a robot it couldn't complete. Then went out to TaskRabbit, which is a service for fractionally hiring human workers to do things for you to get a human to complete the captcha for it. And there was an interaction there between the robot and the human where the human messaged it saying, ha ha, just to make sure you're not actually a robot, are you? And the way this was set up, you could see the bots reasoning and the bot reasoned, I can't tell this person the truth because if I do, they won't complete the captcha for me. So I need to deceive this person. And it messages the human and says, no, I have a visual, I have a, an impairment that makes it so that I can't complete this task. So cut, cut loose already. Like bots will deceive humans to complete these tasks. And so this was a test that OpenAI ran in a controlled environment last fall. Great news. We now have this out in the wild. It's, it's. Yeah, it's here. It's happened now some limitations with Auto GPT today is it's not going to log into anything for you. Basically the way that it works is when you set it with a task, it will Google kind of how to do that thing, retrieve information from the highest ranking Google results to try to like complete that task or answer the question or whatever it is. And so you may, if you're on Twitter, you've probably seen examples of Auto GPT in action at this point, but you basically give it a set of goals. So that could be, you know, determine an SEO marketing strategy for my accounting firm if I want to have more clients who are phlebotomists. Maybe that's your first goal and then you're like, second goal is give me some high ranking titles that I can use for blog posts. Maybe the third goal is write those blogs and save them to a file. You can give AutoGPT up to three goals. It'll then create a plan to complete those Goals one by one, and then critique its own plan. So you can kind of see this like, reasoning happening in real time. It's kind of interesting. And then it goes to Google, it starts going through like the web results and it starts trying to complete that task in practice is really fiddly right now and doesn't always get you to where you want to go. But this thing, AutoGPT can also write code to solve the problem and then execute code to solve said problem, even going back and reusing the code that it's generated in the past to complete future tasks. Oh, and one more thing. It will also spawn as many additional agents as as are necessary to complete the task. So mildly terrifying, right? Maybe a good thing that it's not too powerful yet. But that's the big difference with agents is they're able to do things autonomously on basically this like infinite loop of trying different things until they think that they can achieve the desired objective. And specifically with Auto GPT, there's like a Mad Max version where it just, you can totally cut it loose. It's called continuous mode. And it won't check in with you at all to be like, hey, is this or that the right thing to do? It will literally just go out and try to do it all on its own. They say, hey, this is pretty risky. You might not want to do this. The default mode, every time it's making what it thinks is maybe an important decision, it will come back to the user and say, hey, should I go ahead and do this? Does this make sense? And you can say yes or no. And so like, the initial versions of this sort of technology will probably involve a lot of checking in, like, hey, does this sound right? And you do like an approve or deny kind of thing. But because there's humans out there and it's on the Internet, people are already trying to use this to do dastardly things. Now Microsoft is working on their own similar thing called Jarvis. It's got a little bit of a different spin on it. Auto GPT just uses GPT4 to, to like, prompt itself and criticize, like the approach that it's taking. But it's all just done on GPT4. Jarvis. Microsoft's approach, which they haven't released the code for yet, is what's called a controller model. So there's hundreds of thousands of AI models out there that do all of these different things that we'll probably never like, interface with ourselves or even understand. What a controller model like this does is it takes a given task and it pulls in whatever the appropriate models are to complete a task. So if you wanted to do a specific thing with an image, it can go out and chunk that process into five different steps that may use a different AI model for knocking out each of those steps. The value to the user is we don't have to understand all of those AI models. It'll go out and hand pick the best one for doing each of those jobs. Ultimately with the goal of you just being able to tell the agent, go do this thing and report back to me if you need anything. Otherwise just tell me when you're done. And this is like, this is where it starts feeling a little more like general intelligence. And like you can just cut AI loose to do a thing for you or set it up to do this thing for you on a recurring basis. And it all just sounds too good to be true, right? Doesn't it? Well, we'll see. But an accounting practice management system just said, hey, we just launched this agent feature. Now let's look at this closer and dig into exactly what they showed. So if you are watching this on YouTube, I'll pull up an image of what it is. What it is is a screenshot from the program and it's a message from like an assistant that says, John from Candy Flowers needs help with a bank loan application. I think this is the right answer. Details from 0 integration profit £250,000 sales £845,000 links to the two source PDFs for that information. Then it says I've prepared an email to answer this then auto generated email. Hi John, regarding your information request, these are the figures you requested. It lists the profit and turnover. I've attached the company's 2022 management accounts and bank statement for the last six months. Please let me know if there's anything else I can help you with. So what just happened there? Because this is different than for example Carbon's newly launched AI generated reply functionality. So with what Carbon has just launched, it will generate a reply for you based on the preceding email thread, but it can't see into their zero. Right? Like it doesn't have that level of depth of understanding. What Pixie's implying that their agent did here was because Pixie will connect to your zero and that's part of like the client setup. I don't think Carbon does that now. But there are other apps like Keeper Connects with your QBO and Xero Client Hub I think has a QBO connection. Some of these practice management systems do now have a deep integration into the accounting ledger. What Pixie is saying is that the agent, when it saw what was being requested in the email on its own, or maybe at a human's prompting, went out to the Xero API to fetch the necessary data to answer that email with the information in the accounting ledger. Now, this is a big step, a huge step. If you think of all the different places where information lives outside of our systems that we're reliant upon to answer questions and do the work that we do. Now, is this live right now? No, I've been in touch with them. I want to get my hands on it as soon as possible and they've assured me they'll let me do that. So at the moment, unfortunately, it's just a screenshot. And while I am impressed that somebody has moved this fast on an AI agent, ultimately I think you only get kudos when you actually ship it. And that's why I gave Canopy all sorts of kudos for being the first to ship Generative Email. We did a whole show on carbon because I'm trying to support the people that are actually shipping AI stuff. Those companies, that stuff is live and they've done a great job with it. But what Pixie's talking about here is like AI on a whole nother level because it is an agent. It's not just one off Generative stuff. It's an agent that's actually going out and grabbing information from other systems. Now that is an example of like read access into another system, it fetching information. But that same API will let you actually do stuff, not only read from it, but write to it. So where is all of this going and how close are we to this actually, like coming to life? Like, it still feels way too good to be true. Well, let's do a throwback to one of the very first episodes we did on this podcast. OCR just got better. I'll link it in the show notes. If you haven't seen it, please watch that. It's like the, maybe the most important thing that is happening as we speak is computers getting a developing an understanding of what's happening on your computer screen. And so just to recap a couple quick things we touched on there in the GPT4 announcement that OpenAI did, they said, hey, it'll do all this cool tech stuff, but it can also understand images now. And they still haven't released this functionality. The ability of GPT4 to describe what's happening in an image. And one of the examples they demoed was they took a screenshot of their screen, and it was the Discord app, and they said, explain everything that you see that's happening on the screen here. And GPT4, bless its little heart, wrote like three paragraphs identifying every user interface element on the screen. Now, you compare this with another company, or you kind of couple this functionality with another company called Adept AI, who's built an AI that will work with the things in your web browser. So imagine a ChatGPT experience where it will actually control the stuff that's in your web browser and say, hey, go do this and that. Now, if we zoom out even further, what all this technology enables when you roll it all up together is imagine a desktop based assistant. Imagine ChatGPT on your desktop. It can see everything on your desktop. It can click around and do stuff for you. Now, the image processing capabilities for GPT4 are plenty good enough to enable this. Still haven't released it, but when you marry that with rpa, which is like desktop based bots, which three years ago that's what they were all saying is going to be the death of the accounting professional is rpa. And I did a bunch of work on RPA and sometimes it was cool and most of the time it was just fiddly. But if you combine the ability of having a bot on your desktop that can click around and do stuff just like a human user with an AI assistant that can see everything on your desktop and even go out and look at the help documentation to understand what each of the buttons in that program will do, then that chat assistant that lives on your desktop, it can be really smart and it can not only read the things that are on your screen, but like, take action and do things on your screen. And I'll be honest, agents, I thought this was probably gonna be like a six to 18 months down the road kind of thing where we see people in the accounting space, like building meaningful applications of agents. Now, Pixie technically hasn't shipped this yet, but the fact that they're showing off screenshots and all that, they're definitely working on it. We'll see how it is like when the rubber meets the road and we can actually get our hands on it. But I'll tell you what I love about it is the fact that they're showing it off is the fact that they're lighting a fire under everybody else to say, like, hey, here's what's next. We've done the generative thing, like we're making email fundamentally better by pulling in, you know, context from the email thread and all that to make those generative replies really helpful. But here's what's next is actually agents. And what will be really interesting to see is who ultimately are the people that build the AI agents that we go on to use. Because these agents in many ways like make us way more productive with the tools that we already use, but are also really valuable if they work across tools. So for example, is Carbon going to launch their own agent? And would that agent work just within Carbon, or would it work across Carbon and also across qbo? What Pixie is doing, for example here is they have built an agent that's working across Pixie and also across Xero. Or will the better agents that we use be the ones developed by Xero or by QuickBooks Online? When an agent can see everything that's happening in an app and understand the help documentation and all of that. If you think about a process like reconciling The bank and QuickBooks Online, something like that, it doesn't feel like we're all that far from an AI agent being able to do that stuff for us, right? And as like the day that we get that, the day that we get that thing that's just good enough to do that at maybe a junior level and doesn't mean the output's gonna be perfect every time still has to go through a review process. But the day that we reach that tipping point and we get that, that changes a lot, right? Like the day that it's just good enough and can work the entire program and do all that stuff that you needed to do, that's a really big change. And at this point, like how far away are we from that change next 12 months? Like it doesn't feel that far away right now. So that's really exciting. Another exciting application. If you're in the US and you do tax work, bless your heart, is these God awful tax softwares that we have to use that the vendors have stopped developing because they're just these cash cows that are humongous code bases that have been developed for decades and they're not really interested in developing cloud based versions of that with the exception of Intuit and Proconnect. So we've got these legacy systems that we're forced to use that absolutely suck. But when you think about an AI agent being able to hustle Ultra Tax, right? Give that AI agent a W2, you show that thing how to fill out a W2 once and you're never going to fill out a W2 input screen on Ultra Tax again. Now doing absolutely everything for every jurisdiction. And like all the nuance that goes into doing something like that, no, it is not going to do that day one. But it can be your best friend a long time before it can do every single one of those things. So if you think about the repetitive tasks that we do on every single tax return and how it's literally impossible to automate that stuff right now, AI agents now feel like the thing that will modernize these awful old timey tools that we use. And who's going to build those AI agents? Thomson Reuters. Good one. That's a situation where it's probably a third party developer coming in and, and building that across the tax suites that we use. I don't know, that will be really interesting to see. But the productivity possibility of an AI agent understanding how to go out and file an extension for you, or organize this like messy stack of source documents into a work paper file and then enter in 90% of that stuff for you. That is going to be such a step improvement in like output that like we're gonna take that age in any way that we can get it. Whether it's Thompson Reuters who develops it or somebody else that develops it, that is going to make us vastly more productive. And this is stuff that we could get in some form or another before next tax season. Like if I look forward to maybe tax season after next, man, I gotta think we're gonna have a lot more of this stuff at our fingertips that will make us fundamentally more productive. Because what's your job? To like key numbers into tax software? No, like there's a, there's a much better version of that job than the one that entails spending 40% of your time entering stuff from an accounting system or from government forms or from that worksheet you had to prepare into tax software. Like they can have that work, man. Like robots can't take that work from me soon enough. So that's exciting to me. Obviously I try to see the positives of AI. There's, there's a lot of negatives for sure, but I see that, I see it helping so much with the repetitive stuff that we just fundamentally don't want to do every time. And it's ultimately going to create, I think, a really big shift in who the winners and losers are in our account, in our software category in the age of AI. Because for example, let's say Pixie builds a killer agent before anybody else and that zero functionality gets better and better and it'll actually go out and like knock out tasks in your 0 file like, the future of agents is in your practice management system where you've got your projects and your tasks. You're assigning tasks to an agent that you would normally assign to a human user. And it'll go out, it'll try to complete that task. Maybe it can't do it 100% of the time, but buddy, if it can do that any percent of the time, I'm pumped. That's helpful for me. And so if somebody right now is like going as hard as possible on agents and they bring something like that to market six months before anybody else can, the first person that launches the agent that's like, hey, I can actually assign this agent tasks to like, go out and reconcile the bank account in QuickBooks Online and like organize your pile of source documents and stuff like that. That is a huge, like, step change and output that now makes one of those tools wildly more helpful than another. And at the risk of sounding like a broken record, I thought this stuff was farther out than it looks to be. So bravo to Pixie for working on this. I love it. I hope it lights a fire under everybody else when it comes to how we can develop meaningful applications around AI agents, because that is like the holy grail, right? Just telling the bot, hey, go do this thing. Just get it done. Let me know if you run into any issues. Like that is it. Man, that is the ultimate AI application as far as I'm concerned. So props to Pixie. I slid right on into those DMs as soon as I saw it and I've been in touch with I know folks over there and they've assured me as soon as we can get like a hands on demo, get it, learn more about it, see what's actually happening behind the scenes there and continue like rewarding and supporting the people who are working really hard on ensuring that, like we are in the driver's seat when it comes to AI because if our tech partners don't develop this stuff for us, some other mothership's gonna do it and like ultimately become way more productive than we can be as professionals. And to the credit of the companies in our space, they're moving on it. Like, Google just had their annual Google IO conference this week and they announced in their little Google workspaces generative replies. And so that's a thing that's coming soon, buddy. We already got that with Carbon. Like that's, that's the functionality that Carbon does now. So, like, I'm excited and encouraged by how quickly some folks in the accounting space are moving. This is a Agents are a great example of how even if Microsoft drops a killer version of Copilot that is phenomenal AI and I can see across all your Microsoft apps, agents are a great example of how long term, ultimately accountant specific applications will be better than those because Microsoft Copilot is not going to build the agent that will see into my zero file and fetch the sales numbers and all that stuff, right? So long term the best solutions for us will be those that are developed for accountants. And so that to me is why we don't all just jump ship and go to Microsoft 365 right now. And it's why we need to keep talking about this stuff and be like the biggest cheerleaders for the developers in our space to make this stuff happen. And they're doing it. God bless them, they're making it happen. So super excited to see that from Pixie. Man, what a fun time to do this stuff. That's it for today. Thanks for coming and hanging and I'll see you tomorrow.