227 How AI Is About To Make Accounting Quality Assurance Easier

I try not to get into the habit of speculating on what AI is going to do for us because there are people with way bigger brains than, than old smooth brain here that can't seem to agree on what AI is going to do in the next three years. But I can tell you something that is seemingly, seemingly going to be here in the next 12 months that is going to represent a massive change, would have completely impacted how I came up through the profession. And that is the basic ability of chat assistants, agents, whatever you want to call them, to do like basic quality assurance work for us. Really easy to set up really practical ways. I've been doing a lot of talks lately and this is one thing that keeps coming back around as like, yeah, no, it all things seem to point to us crossing a threshold in the next 12 months that, like, changes a lot of how we do, like technical reviews, that sort of thing. So running through some examples of what that looks like today, why it sure seems like something that's going to be wildly different 12 months from now and what that means for now, like, do you prepare for that? You just dig your nails into the, into the armchair, the sofa, and just get excited for it. We're gonna talk about it. Come on in, buds, it's Jason Daly. All right, this is a fun one. This is only reason I want to share this is it's great news. It's one of those things that cannot come fast enough that we should all be super looking forward to. You may have seen a wee, wee GPT that I put together a few weeks back in my weekly newsletter. Newsletter, JSON cpa. If you know, you know. And basically what this GPT does. A GPT is like a kind of a reusable chatgpt conversation almost where you can customize it with specific instructions and upload information to it so it like, knows stuff. You create this GPT, it's like this custom assistant and then you can share it with team members, share it with other people externally and all that. We've explored a bunch of cool use cases for GPTs on the YouTube channel. That's YouTube. Com. Well, the thinking behind this, this GPT is called a Balance Sheet Checklist Completer. And I don't know how much most of y' all still rely on checklists, but I could tell you they were, they were the hotness. When I was coming of age in this profession, we had. Who are all the ppc. We had all these checklists from ppc. If you know, know all these different things. The first firm I Worked at this guy had a. What did he tell me? He's like, if you're gonna, if you're gonna do a different process, name it and then you can charge for it. He had like some sort of catchy little business wisdom around the fact that he had this like 20 some point process. And all it was was like a checklist that went under the lead sheet of things that you sign off as having checked and all like very basic trivial stuff. But yet checklists. I mean, as much as I hate to say it because of how annoyingly tedious they are, checklists, super, super helpful because a lot of the aspects of what we're doing, month end closes, reviewing tax returns, all these things, like, it is, it is repetitive stuff. It's different clients, different circumstances. But the, much of the checking of the work, the, let's say the detail review, not necessarily the technical review, but the detail review is pretty rote, like repetitive stuff. Particularly when you're, when you're growing your firm and you're trying to standardize things more and you need Jim to be saving and annotating and naming the files the same way that Tina does. And why does nobody else see that this is as much of a problem as me, that you're not doing it the same way? And you tell them, you know, nobody else is going to be able to follow your work. And they're like, that's not my problem. Until they get like three years down the road and their job becomes to like upskill to other people on how to do this stuff. And then they're like, oh, okay, now I get it. Nobody's interested in documenting their work until it comes time to actually show somebody else how to do it and follow their work. And so there's all these things that it would be like, boy, wouldn't it be great if something could just come through and like automatically make sure that this stuff's all happening the right way. You know those SOPs that you build. Imagine if you had an SOP for, you know, we had an SOP, for example, for like the naming convention for all the files that get saved for a month end close. We have, we had an SOP for the, the like folder structure of what needs to be there for every single month, then close, so that somebody could go through two fairly concise docs, didn't want it to be too long, and understand exactly all the files that should be there, and act exactly how all the files should be named. Another one on like annotating and marking up work papers, what the expectation was there and you know, the first couple times somebody did a month end close, they got it. They're like, okay, I know the expectation. The reality is did they follow that stuff to a T every single time? Did they forget things sometime? Absolutely. And unfortunately it ends up meaning that a lot of the review function and we talked about this actually back in the quality assurance episode, which if you haven't heard that one, I highly recommend it. It's a, it's, it's a banger, it slaps, is a good one. Maybe if you're on the YouTube feed, I'll link it at the end. But a lot of quality assurance isn't like a big brain thing. It's like, oops, did we save the right file here? Is it named the right way? Did we not do anything that's obviously wrong? Like just happens with the volume of the stuff that we do, it happens. Unfortunately, a lot of the work that we do is just, it's high stakes. I don't care who you are, once you, when you do something a thousand times like you're, you are bound to make a mistake. And in some of the work that we do, those mistakes can be really expensive. And so checklists just like, it feels like they've always been around and there's always been this impossible tap dance between, you know, the checklist could really be a thousand things long. But it's like, well, nobody's going to do your thousand thing checklist. And so you give people a checklist is 20 things long. And it's like nobody's actually reading your 20 thing checklist. You do it once or twice and then by the third time you're just like skimming through that thing, trying to, trying to crush it as fast as you can. So my thinking was what if we got AI to complete a checklist for us? And the reality of where AI is at today is it's a long, long ways from being able to feed it like a long checklist and say do all these things on a checklist. It's much, much closer and in many cases good enough to give it a single thing on that checklist. So let's say you've got a 20 point checklist before any month end close rolls out of the shop. In most cases it can do a single thing actually really, really well. But the more, and I say it, I'm talking about AI. So this could be chat GPT, literally just racking up to chat GPT with a balance sheet and a P L. Any other like related work Papers that would be required to tick the box on this ch list item tell Chat GPT. Here's the thing that you're supposed to check, and it does it pretty darn well. So the GPT that I had made in the newsletter, I had it check three very basic things. First, it checked that total assets equaled total liabilities and equity. I know, knock rocket surgery, but don't act like you haven't gotten it wrong before. This is a safe space. I haven't. I'm talking about you. Second thing it does is confirms that the reports run a cash basis that was the default for us. Everything was cash basis. Third was that it was comparative previous current period of the previous period. So all of our, all of our cast stuff, our, our bookkeeping side of our business, all reports cash basis comparative. So it checks those two things and then it checks that total assets equals total liabilities and equity. Because somehow in 2024, you can still be fooled by the software you're using into thinking that that's a thing that it would just do. Yet it does not always do that. So that's the instruction to this, this GPT. And it's, it's. I say GPT. It's the same as going to ChatGPT and instructing it. Here's task one, here's task two, here's task three. And telling it what you want it to do. And so, for example, I say prove whether total assets equal total liabilities plus equity. I give it some instructions for how to do that. So for each of these, as I was testing it, I found that if I gave it a little more granular directions as to how to figure this out, it performed a lot better. And I was just leaving less for it to kind of read between the lines. Confirm the financial statements were provided and run on a cash basis is primarily done by looking for a cash basis label in the report and confirming no accrual basis label is present. The same way that I would explain this to somebody who's totally green. So it's basically like the explanation that you would give to an intern who's like, I kind of know what that is. And that's kind of where ChatGPT is at right now. Like, it kind of knows what that is, but it doesn't always know how to apply that knowledge. So in the case of this little GPT does those three tasks, there's actually a fourth task where it creates a work paper because ultimately it has to prove its work just like any other human that you work with. It has to leave a breadcrumb trail showing. Here's my analysis and why I came to the conclusion that I did. Cool thing about ChatGPT right now is it can actually annotate and create and annotate PDFs. So what it does is I give it a balance sheet and it creates a cover page for that balance sheet that summarizes the results of each test. Each of those blocks box, each of the tests are in a different color. And then the second page is the balance sheet that I gave it. And it highlights on the balance sheet the color of all the relevant details from the balance sheet to each test. So if test one is blue, it'll highlight everything in blue on the balance sheet that was relevant to it passing or failing that test. Pretty darn cool, right? But like, not unlike a human would do. And it did all that because I told it to annotate it that way. So you can make it annotate things any way that you want it to. And it works about two thirds of the time and about a third of the time it just kind of falls on its face and you have to tell it to try again. And usually after a couple of tries it gets through it. I don't think that I have seen any like false positives. What would that be? False negatives? I don't think that I've seen it give me any wrong answers. Certainly not to say that it could. More often what it does is just like falls over midway through and I have to tell it to get itself back up and try again. Whole process takes probably less than 20 seconds. And that is where it's GPT4. Behind the scenes is the language model powering Chat GPT. That is where language models are at today. That's kind of how far you could take it. So could you do a big impressive version of this right now? Honestly, probably. It would probably mean, you know, like standing up a little product to where it's a totally separate, you know, conversation with GPT. Each checkbox item that is trying to complete, like, you probably just make it razor focused on, you know, a standalone conversation and process to tick each of these checkboxes. And that may feel like a lot of effort for something that's otherwise trivial and not that hard to do right now. But the thing with language models is each new generation of language model, generally what the previous version was capable or what was like the ragged edge for the previous version is absolutely table stakes for the next version. Easy peasy. We're talking on reasoning, so ability for it to understand we're talking on speed, we're talking on cost. We now have open source language models that are arguably every bit as good as GPT4. People are already getting. This is, this is the new language model from, from Meta, from Facebook. People are already getting that open source Langu model to generate at unbelievable speeds, like for it to work really, really quickly, actually generating like over a page of text per second much, much faster than anything ChatGPT does. And that's only the third version of Llama of Meta's language model. And so overnight, something that was really novel is now table stakes and can happen near instantly at virtually no cost. And so it requires you to kind of step back a little bit and be like, okay, what are the things that it seems like AI is kind of close to being able to do right now? Well, it's gotten pretty darn good at research. Like if we're talking about using cloud or something like that and giving it some, some technical resources, like you know, 500 pages of stuff and say, you know, pull out all the things in this body of text that are relevant to this thing that I'm researching, 1099 or foreign income in this specific situation. Something like that actually does that really darn well right now. But stuff like this, this kind of like checklist stuff where it's kind of on the edge. This episode is sponsored in part by Copilot. You ever struggle with client portals that just, they're just not quite what you need? You know, they like, they don't have all the stuff you want or you want to tailor them or even want it to look different for client A than it does for for client B. Check out Copilot. It is like a hyper flexible portal platform. They call themselves actually a platform rather than just an app because what they want is for you to build the portal of your dreams on top of it. Super customizable. Even change what it looks like by client or by client group. And they recently launched an app store. So like custom integrations, like stuff built on top of the portal itself. They got a whole library now, stuff that you can pull into it. Or actually hired a developer to build something like completely from scratch to augment your client portal experience. How cool is that? Like theoretically you can build just about anything into there, connect it with like third party apps. Copilot's got a great API, so you can even wire up some of your own stuff with, you know, zapier, make that sort of thing. Because we all know what it looks like to have like a disjointed client portal experience or ask them to jump around to a bunch of different places. Nobody likes that. It makes you look bad. Think about it. You want to level up that client portal? Check out Copilot link in the show Notes this episode is brought to you in part by Team up, helping you recruit top Filipino accountants without any ongoing monthly fees. The difference between Team up and all the other offshoring options. The Team up helps you hire staff directly. No middleman. You work directly with your new hire in the Philippines. Hire the person, not the company. Guys, Gals, gang. Here's just a few reasons to hire directly. You have access to higher level talent. Makes sense. You have complete control over team culture and training. They keep 100% of what you pay them and it's a lot more affordable for you. So you can retain your team for the long term. Team up could source accountants with experience working at US or Australian firms familiar with tools like Zero, QBO and DeX. Also recruit specialist roles, team leaders, tax specialists, administrative assistants Thought experiment what if you had an executive assistant for the first time this tax season just just throwing it out there. What would they do? Start at that email video I did on the main channel recently. Get help with that stanky ol inbox. I digress. Team up recruits these talented folks for a flat one time fee of 4,000 US American dollars. That's it. 4K one time. Somebody at Robert Half just did a spit take. Robert Half reference. We ever going to get Robert Half to sponsor this pod? Not anymore. And they can connect you with an affordable employer of record if you need help with payroll and compliance once you hire that person. Big fan of hiring in the Philippines. You know I did a bunch of that. Check out the link in the description to learn more about Team Up. A day is going to come when a new language New language model from OpenAI drops. And a lot of people think that could be any time. They've said it will definitely be this year. There's mounting pressure on them to ship something because GPT4 is now far from the best model out there anymore. But a day will come where it's not possible and then the next day it's all possible. And so I'm interested in exploring what does it look like to have language models that can complete checklists for accounting firms? Because I don't. I don't think it's outlandish to say when we get that next model. In all likelihood, I don't know that we're ever going to have to complete a Checklist again. I don't know that you're ever going to ask a human to complete a checklist again. Right? Why would you, if the AI can do that for you? If I'm a manager, I would sooner write an SOP with instructions of how to complete a checklist item and hand that off to an AI that can run instantly, where an infinite number of those agents can run concurrently at the same time. Right? Like. Like I would sooner hand that off to an AI than I would an intern. In fact, at that point, it's almost insulting to put a checklist in from, in front of a human. And to be fair, right now it maybe already is a little bit insulting to put a checklist in front of, in front of a person. And I'd actually say oftentimes we don't just because we're like, we're almost like filled with shame at how manual and kind of dumb checklists are. But what is the alternative? Like, there's also tremendous value in process. You know, we've talked about Boeing and the 737 stuff and the absurd level of control that goes into constructing an airplane and how there's an entire approval process to remove a bolt or screw or anything like that, and just how locked down those processes have to be just on a. In a completely other universe that, like, I cannot get my head around. But when it comes to making the software sausage, man, there needs to be standardization there. Whether it's a checklist everybody's running through, whether it's, it's SOPs, which I honestly are, are really just a, a more trendy version of a checklist. Like they're the same thing at the end of the day. But as soon as I can take that sop. And maybe an SOP is actually a better example because it's a little more narrative and gives more description of, of how to do it. If I can check that SOP or a checklist into a language model, this could be chat GPT, it could be a custom GPT like we were talking about. It could be an app, it could be some, you know, creates an app for accountants that is just on the back end. All it's doing is via the API going to GPT, and the app is just, you know, simplifying that process for you, maybe. But on the back end, like the, the AI sorcery that's happening there is the same. Once we've got all this, I, I realistically don't know that, that a human's ever going to go through that list. Again. And so it's, I'm, I'm in love with this notion of like basically self completing checklists. So if you think about in the past where folks have given out checklists for how to standardize procedures and all that, and that's, that's useful and I've definitely relied on those in my firm in the past, especially in like financial statement preparation. Right. You see a lot of that stuff. But I think we're in the very near future, it's not going to be enough to just give away a checklist. I think you almost have to give away the checklist that will complete itself. Like, that's all people are going to use. Like, cool, you got this checklist, give me the version that will go through and review that stuff for me. So if you think about what that could look like a, you know, I use the example of a balance sheet checklist complete, or let's say you've got a checklist for some sort of due diligence going through a balance sheet. Could be, balance sheet could be scheduled from a tax return, could be, could be anything. In order to complete those checklist items, assume this system doesn't have access to every file in your system and all your emails and all that. Right. Because probably won't. For a given checklist item, what are the requirements? What information would it need to be able to sign off on that checklist? For my balance sheet checklist completer, it's probably the balance sheet, probably all of the work papers associated with the balance sheet accounts. And so for a given checklist that will check these things, you have to provide it with the balance sheet, you know, supporting work papers for the balance sheet, accountants, all that stuff, you upload that to the checklist and then the AI agent, the language model, goes through those checklist items one by one and gives you a work paper that says, here's all the work that I did in my findings. And right now that kind of works within 12 months. That's table stakes and near instant. How cool is that? How exciting is that? All of a sudden it's like, ooh, oh man. Like checklists are about to be super cool. Because that thousand step version of the checklist where you're like, ain't nobody got time for that. If an AI could do it for you, instantly you'll be like, oh no, actually come on in, that would be great if you could do that for me. And so then all it takes is like one person building that, you know, thousand item version of the checklist. If you Think about tax prep. Maybe that's, you know, a checklist for every schedule in a tax return. Oh my gosh, we're just talking about an entirely different level of like comprehensive review. And again, I'm thinking about this more through the lens of like detail review, not necessarily technical review or I'm asking it to do big brain stuff. And then maybe that's why I keep coming back to checklists is if it's something that I can give really explicit instructions for how to do. And there, I mean, honestly a massive, the majority of the work that we do is this. A huge percentage of the work we do is this. It is the king of the numbers into the software. It's making sure that this number on this thing matched this thing that where it came out here, that stuff that can be boiled down to an SOP or a checklist can be done, you know, near instantly. And we've had quality assurance tools in the past and there are, there's lots of quality assurance tools out there that folks use and are totally worthwhile. So what's the difference now? And why would I not just sit back and wait until you know, one of them make a really compelling version of this? The main difference I'd argue is while like off the shelf software will be able to do this at scale in a locked down way, it's not going to be flexible to you. Most of us do not know how to program. That's not a thing that you know how to build from scratch. Right? But the opportunity here, the exciting thing here is this quality assurance is working based on semantic explanation based on you telling it what to do. And for as much as I've struggled with the whole notion that AI is going to like replace software user interfaces and we're just going to talk with computers now and that's it. And we won't even have like reports that we look at or anything like that. I'm like, man, it doesn't make sense for accounting. What does make sense for accounting is in a situation like this, I don't know how to write code for it to go and figure out if a balance sheet as cash basis. But I can say a paragraph that is like how I would explain to an intern how to go check that it's cash basis. And the wild thing is that's good enough. Isn't that cool? So yes, we can sit back and wait for like people to ship really cool QA stuff and that will actually be a thing. Frankly, it'll probably be built into a lot of the Tools that we use. The biggest problem, as with almost all accounting tech, will be the fact that the information to do that due diligence is fragmented in different systems. So QuickBooks can't see your work papers. So QuickBooks won't be the one to do the checklist for you. Most of your practice management systems still don't integrate into the ledger. So they may have the work papers, but they don't see into the accounting ledger. Somebody still has to get all that stuff into one place to make it really useful. But for me, the operator, I'm still going to have SOPs. I'm still going to have stuff that falls outside of just whatever that standardized solution looks like, right where that, that is like that standardized solution is now like ground level. That's now what everybody's working with. We now all have that capability. But the stuff that falls outside of that, you know, for my beekeeping clients that have, you know, unique needs or the ways that our firm does things differently, and there still will be a lot of things that you do differently. Like, it's not as if accounting ledgers, computerized accounting ledgers, led to us all doing accounting the same way. Like that didn't happen. AI making these tools better is not going to lead to us all doing accounting the same way. But the beauty of AI in our hands, it'll make our developers better. But in our hands, it means we can basically program it just. Did I say program? We can program it just by telling it what we want to do the same way we would a human. And that, that I can get behind. When Bill Gates says like, oh, we won't have interfaces or buttons or anything like that, I'm like, I don't get that. But what I've come around to is the exciting part of that, the exciting part of the notion of just all work happening in a chat box or over voice or something like that. What I am seeing more now is, oh, you know what that actually looks a lot like is working with other human beings. I mean that is how we communicate with other people and instruct them on what to do. And the near future looks a little bit like talking with language models and AI agents the same way that you would a human being and then asking them to go reason through these things and complete these, these steps for you. Big question obviously is, is the agent good enough to understand what it's supposed to do from how you explain it? And this is where like software companies will, will be able to build like a very robust version of the standardized things because they'll ask the questions just the right way. They'll use the perfect language model for, for that specific problem and all of that. So for the standardized stuff, like they'll squeeze everything they can out of it. But are the language models going to be good enough just to understand it? Honestly, I would argue today GPT4 is probably as good as an intern, if not better. And whatever the next model is, 4:55, whatever it is, it is, it's. Sam Altman says it should be as big of a step as each of the previous models before. And every single step has been a massive step in reasoning ability. So we have not seen anything thus far that would make us think that it would not be more than capable of doing these sort of, you know, checklist sort of due diligence tasks. And that's if, if they're just going in blind, if you don't even give them any context. I mean, cool things that GPTs, for example, and ChatGPT can do is you can give them custom knowledge, you can upload, you know, supporting documentation to them. So let's say you've got, let's say you did a, like a, like a training over zoom of how to do this whole thing with your team and you save the transcripts. Chuck that in there, there now that is going to inform how it works. The next GPT model is said to be multimodal, which basically means you can give it inputs of different types, text inputs, picture inputs, video inputs, a video of somebody actually going through and doing it. Like you can provide that. Like that's going to make it more accurate. And the wild thing I always try to remind people about AI is that anytime we make a tech decision in our accounting firm, we are hitching our wagon not only based on where it is today, but where we think that this partner can take us. Like equally important to what it is today is are they going to continue developing, listening to our feedback and keep making it better. And there is nothing on the planet that has ever grown as fast as AI is growing right now, like in the quality of what it does. And I get that, like it's easy to be impatient when it's like, yeah, but it's still not here yet. Still like not at this threshold quite yet. But it is worth pausing and consider what changes the moment that it is. And this is one of those examples where like as soon as it's over the threshold of being able to complete a checklist by itself and nobody ever has to do that again, well then like what's next after that, is it actually completing the checklist? What if it could just like fix any issues that it saw? Right? Like, so then do I actually have to complete the whole thing or does it go through that checklist and then just kind of like put those things together based on what would like, make it meet that criteria? And before you know it, like, you've got this thing doing like actual meaningful parts of like the technical work now that's much more abstract and harder to get our heads around. So like the idea of it just doing quality assurance, that's. That's maybe a little more palatable. This episode is sponsored in part by firm360. You ever been big boyed by a software vendor? Like, you're like, hey, I really need this, this thing, or this thing's totally broken and I need help. And they're like, let me, let me shuffle you over to here, let me get back to you on that. And it eventually becomes very clear that they have either no interest in helping you or disagree entirely with everything that you're saying and are not going to do a thing about it. This becomes more of a problem. In my experience, the bigger a company gets the old like, not my problem sort of thing, man, does that get frustrating? Especially, especially when it comes to like your practice management system that is the core of your accounting firm. Like, to feel like I've invested all of this stuff and we've like dumped all of our company data into this tool to now just like get the shaft like that. There is nothing worse than getting burned by your PM provider. And man, it happens. Like, you see the people who have had bad PM experiences. Firm 360. I know a good number of folks in the wild that use firm 360. And it's admittedly really hard to cut through the noise on practice management systems to know like, oh, what's good at this? What's good with that? The one thing you consistently hear about Firm360, phenomenal support even in there, like G2Capterror reviews in the past, I've read some of those for ad reads. And the common thread that you see is like a next level support. So if you are overwhelmed with PM options, frankly, five years ago there were very few Cloud PMs that were worth using. Today there are so many more. If you're in a spot where you're feeling like, man, I'm just overwhelmed with all this, but at the end of the day, you want to make sure they got your back, man, they're gonna be There for you with whatever needs you have. Having a rough day, a warm glass of milk at the end of the night. You never know. Listen, all I'm saying, Support is important. Firm360 is good at it. That's not even in the ad re. That's what people tell me. So if you're looking for a practice management system that cares about you as much as you care about what you do. Burn360Link in the show notes. Was that a bad ad read? You get the, you get the idea. Big, big question will be well, how am I ever going to trust it right? Like how do I ever trust the output of that? How did that actually save me any time if I'm just going to go half back through and check the whole thing again? Welcome to working with humans also right. Like when's the last time an accountant trusted anyone on this planet or anything that they were given? That's how it always is. But this is like a worker that is available 247 costs next to nothing. You can run as many of them concurrently as you want. I mean you could, you could have a thousand of these processes kick off at the same time. So anything that you can do with an agent you can do in a, in a virtually limitless way. And so are we going to blindly trust the output? We sure won't. In the same way that you're not going to trust the output of an intern of, of one of your colleagues. It is just something really powerful that we're about to be plugged in that we're about to be able to plug into the machine. And I don't know what roles toes it steps on. Like is that stepping on the, on the role of the intern, is it stepping on the role of the technical reviewer? I don't really know to be honest. I think we're just going to kind of have to see when we get. But it's definitely something to start thinking about and a common trend. It seems like every time we talk about AI on this show is how we always come back to. That's really cool. Just make sure you've got your, your documentation spiffed up and it's all ready to go right like that the AI is only going to be as good as what you can instruct it to do. And most firms these days, they don't. They still don't have a quality assurance process like they are. They're reviewing things based on hard earned expertise and just knowing right when they see it, you know, and that's about it as Opposed to trying to back into, like, how do we systematize right? And how do we make an organizational rule for what is right and what isn't right? Because a machine can follow rules. And like, people still talk about AI like it's this black box thing. It'll for sure be able to do cool, novel things. I just don't know how useful that'll be in the context of how we work. What I need is for that thing to be able to follow instructions perfectly and then report back clearly on what worked and what didn't work. What I don't need is something to just like be cut loose in my file system and be like, figure it out, you'll do great. Like, that doesn't actually get me anywhere. If anything, it maybe goes backwards, but a super clear SOP of like, here's all the things that I need you to do. And it goes and it does all of those things like near instantly. Oh, buddy. That is extremely exciting for me and our trust. And it's all going to boil down to the like, the quality of that work paper in the end. Right. And I can tell you the biggest complaint I always got from staff when we would tell them that their work papers weren't up to snuff. The complaint was always, but that's just going to take me so much time to make a better work paper with AI that constraints like, poof, goes out the window. What is the perfect, most clear workpaper? Time is no object. Like mark it up to the end of time, make it just an absolute work of art. Like, AI will do that for you instantly. So like, as we get there, and that's really exciting, will we stumble into new shortcomings and like new things? It can't do that we didn't expect almost certainly. But boy, is it exciting to think that we could be like mere months from this bottomless assistant that can just do your bidding based on like whatever you describe to it at the drop of a half. That that's super exciting. On the, on the integration problem, the fact that we have information in all these disparate systems, like always been a big problem for any accountant, but particularly in public accounting, especially if you do tax compliance stuff. Those systems just don't talk to each other. OpenAI is said to be launching a desktop based AI agent solution this year. So what we're talking about is something that could literally operate across your entire desktop, see everything that's going on, jump across different apps. That's what's so fascinating about what's happening with AI right now and, and quote, unquote agents, which are basically just like a language model kind of being cut loose to do a task. Like given access, given the ability to click and type and, and make API calls and stuff like that. The fascinating thing about agents is everybody's working on them. Like, everybody's working on them. But am I going to use Intuit's agents? No, because the agent that I want is the one that'll do Intuit stuff, it'll do QuickBooks stuff, but it'll also work with all my other apps and see into my email and do all of these things, like across everything. So who builds the agent today? Like, well, Intuit's not going to build something that's going to like work nicely across all my other apps. Maybe it's OpenAI, like this new kind of third party. Like it's the quote unquote AI company that creates that agent. I was at a conference in the UK last week and talked with loads of developers who were still in stealth and all working on various AI agent solutions. And it's really exciting because the incumbent companies with the big successful platforms, they can only make, they're only going to make agents that do stuff on their own platform. But the most useful agents for you and I are going to be the ones that are doing work across all those systems. There's almost an entirely new category of tech vendor that's like giving us AI tools that work across all the tools that we already use. And almost without exception, like technologists, like software folks think this is where things are going. Over the next couple of years you'll have these super powerful agents that are using your apps for you. And it just, I mean, it kicks off a ton of other interesting things. Like if that actually happens, like overnight, you're going to have probably more agent users of software than human users of software. And if you're on a software platform, yikes. What does that mean for me? Well, usage based pricing or, well, user based pricing just went out the window. I could have a single user with an astronomical number of agents doing stuff all across my platform and there's like little AI agents just doing all of this work, but I'm only charging for like one headcount. So software systems probably have to go from user based pricing to usage based pricing or how much the platform is being used. Also, what does it look like to optimize a piece of software for an agent to use it versus a human to use it? Are those two different things? Are they the same thing? I Don't know, man. But as soon as I can get an agent to like complete a checklist, hop out, double check a thing in QuickBooks, complete a task in my project management system as soon as I can get it to do that, arguably we may already be there. Like, we've. I've talked about hyperrite in the past. Hyperrite.com, it's an AI agent that works in your browser. Right now you can tell it to do stuff. Like right now I can log into QBO and tell hyperrite to go run me a cash basis balance sheet and profit and loss report for last year and it will navigate to reports and it will navigate to balance sheet and it will change the date and it will change the reporting to cash basis and it will click run report and it will say save as PDF and then it will go back to reports for the other report that I asked it to get and it will go through and it will do all those things flawlessly. Already it's a little slow. Takes like three or four minutes to do that. Little resource intensive. Like if I were to do that on a huge scale. But when I have an agent that'll work across all of my apps, all of a sudden I really want apps that play nicely with my agent, right? Like, okay, we've got all these practice management systems, but which one will my agent be able to navigate the best to like complete things for me, right? Man, like, this is going to get so weird and we're so close to this. Like, this is all like, stuff kicking off in the next 12 months. So that's all really exciting and honestly, really unknowable what that reality even looks like in the shorter term. What I'm really excited about, down with checklists or maybe up with checklists. Maybe this is like the golden age for checklists because now we have like unlimited checklist minders. I don't know if you've ever experienced this talking with an AI chatbot. When you give it a stupid big task and you almost feel bad asking it, or you ask it and you expect it to say no or complain. Like when I'll have chatgpt generate like 50 hyper, like micro niches for accounting firms, I'll be like, generate 50. And then I'll be like, ah, give me 50 more. And I'm like, sorry, dude, I need, I need more here. And it's like, happy to help. Bless its little heart, man. That's the thing that's going to be hustling checklists for us and doing all this Due diligence and. And that's just really exciting. I don't know that. It ultimately changes a ton about how the work gets done, at least in the short term. Like, no part of that automates the preparation of things, at least not in the very beginning. But it sure does sound like a really nice thing to have in your back pocket. Right. And again, all this is taking us down the path of, like, how documented are we? Do we have standardization? Because oftentimes in a firm, you try, but it kind of dies because people can't agree on doing it a standardized way. It's a great time to be having those really hard conversations right now, getting the humans to align on a unified approach to doing things, because without that, you can't come to any tool with an expectation of it being able to help you when you don't even know exactly what it is that you need. Right? Oh, baby, did I deal with that big time. Coming into a legacy firm with a whole bunch of people that all had their own ways of doing things like, that is a. That is a big lift. But honestly, never been a better reason to get all that stuff together than the fact that we've got tools on the horizon that can, like, follow a standardized procedure for us. They're going to be a massive help. Very excited for that. What's today? Thursday. Got Q and A. Friday. Tomorrow? Got any questions? Drop them down below. I admittedly usually can't get the Thursday episode questions in time before I have to shoot the Friday one. So usually the cutoff is like, questions through Tuesday or Wednesday. If I miss them, though, I always circle back to them the next Friday. That's enough of my AI ranting for this week. I feel like I had made it, like, a good couple of weeks without going on a big AI rant. It was a heck of a run. Listen, I just want to make sure everybody on the team's prepared. And we're not. We're not surprised by anything. Okay? This crew tuning into this pod is genuinely the most equipped accountants in the profession in regards to all things AI. I can say that with complete confidence. Not to say that everybody's got, like, their AI bandolier, all these different subscriptions to various AI assistants, all that stuff you gotta be dipping a toe in for sure. But nobody's going out of business today by, like, like, not spending half their day in chat GPT. We're not there. Maybe someday we will be. But I can tell you the people tuning into this pod will never be surprised by a thing. And when that thing comes, you'll be like, oh, yeah, we had kind of seen that coming. And that's the best place to be. Just don't want to be surprised. And when you got all the information, you can decide for yourself, like, when is the right time to dive in and how deep. Like top of my tutti toes like, or neck deep. Am I going in all the way. Thanks for coming and hanging today. We'll see you tomorrow, Sam.