506 Today’s Fastest Growing Accounting AI: Is It Secure?

Very exciting gang. There is an AI that is vastly better than any other AI we've had to date for doing accounting stuff, for like doing real like meaty work. And it is Agent mode inside of ChatGPT came out over a month ago and people are still not talking about it enough. But we recently published a video on my main YouTube channel. Over 50,000 people watched it in the first two weeks. And it's because the examples we served up in that video, they are so cool and think this is the first time like accountants are kind of riding the wave of like, okay, I can actually see how agents are this incredibly transformative thing that is just around the corner. But as you may suspect, there are a hundred security questions. Right? And that's kind of the premise of the video is in some ways that Chat GPT agent functionality, it's problematic from a security standpoint in some new ways that we have to be aware of. So how do we navigate that in the course of running our firms while giving our teams the boundaries that to use this stuff that's really cool and really useful, that's what we're going to go deeper on today? What are those boundaries and how do we confidently address the security questions around this? To know that we're doing the right thing by using it, but also to have confident answers as we're discussing it with clients, with our teams and all that? Let's get into it. This video, it's one of the biggest videos we've ever done, like highest performing. And anytime that happens, it finds a new audience. And so we got a whole ton of people into the video that are like, wait a minute, since when can you use AI and accounting stuff? Since when can you use AI with sensitive client information? So there's a lot of folks finding the content for the first time, asking questions that are like totally valid and responsible questions to ask. And if you've been around the podcast for a while, you know that there are completely secure ways to use AI. The fact that there remain non secure ways to use AI does not mean that there aren't completely secure ways to use AI. But with ChatGPT agent mode, that is this sort of new functionality in ChatGPT that can like navigate websites for you. There's some new risks with that. And the biggest one to be aware of is If I log ChatGPT agent into a service that I use and then tell it to go do this thing and then run off and make a burrito, do I trust that what it does won't be problematic? Won't create issues. And we're talking about this today. Can I trust an agent to go out and do this thing correctly? This may genuinely be the biggest technology talking point of the next decade. I told a robot to go do that thing and it messed up, or can I trust it to do that reliably or something like that? This problem is gonna get worse before it's gonna get better because these agents are getting more and more capable all the time. But it's not unlike self driving cars in some ways where you can see a world where like self driving technology, there's actually fewer accidents than when humans drive. But what about when self driving has to deal with dumb things other humans are doing or you know, when it has to make a hard decision? That's a much bigger headline than maybe the human version of the same thing that happens every day. I mean, goodness sakes, we, we are all error prone. I mean, I mean, not us, I'm not, but all the people we work with are error prone. And so this whole question of cutting an AI loose to do something without our oversight, there's people that will struggle to do any version of this for years. And honestly, it's the same type of people who say this one wind up, this one's going to hurt. The same type of people who say I'm never going to delegate this because it would take me more time than it's worth to do so. Working with AI agents I think is not going to be all that much different than or working with other humans where you're delegating things to it. It's going to find new and creative ways of doing it wrong that you hadn't expected, just like humans do. And at the end of the day, it is a learned skill how to work with this thing effectively in the same way that working with other humans effectively is a learned skill. And the sort of like QA process, how you ensure that the work product is correct in the end, that still boils down to you in the same way that cutting an intern loose in a QuickBooks file and then breaking everything creating creates problems for you as well, right? And if you've played with ChatGPT agent mode much yet, and I super encourage you to do it, what makes it amazing for accounting stuff is it is very, very good with spreadsheets, like nearing human levels, like average human levels of capability with spreadsheets. And you and I are not average with spreadsheets. We are exceptional. But this bottomless, unlimited AI agent that can just slog away on this stuff like for infinite amounts of time. Like several of them working at the same time. That's really valuable in that video. There was one person who listened to the original podcast we did on ChatGPT agent who shared on LinkedIn they had like a big meaty annual or sorry monthly spreadsheet they have to put together for a client takes a half day or something like that. They cut ChatGPT agent loose with it and it did, they said as good or better than junior accountant level work on it. They still had to finalize it, but that's what they did before anyways. They would give it to a junior to work it up, they would finalize it, then send it off to the client. That is today, gang. That is where it is at today. And 12 months from now, we're going to look back on where it is today and be like, man, that sucked. So this whole like tipping point of agents being good enough to do xyz, we are like already teetering on it and when we're on the other side of it, it is going to change so much of how we work. When an agent can confidently navigate the different tools that we use and do like real accounting work. And that is, that is exciting to me. But let's address the security spookies and what I've done is I have pulled every security spooky related comment from that video and some of them have easy answers, others of them don't. But we need to develop the muscle of being able to confidently speak to. Here's why this is totally fine and we're using this responsibly because most people, the majority of people don't understand still that there are completely secure ways to use AI in supporting our clients. And we've talked about this on the podcast before. But for every fear you have of like, do I have to disclose the usage to a client, what would my clients say? There are just as many clients for whom you working with AI, that is a feather in your, your accountant's cap, your accountant's visor. Because if I'm hiring, you know, a marketing agency to work with me right now and I ask them in the interview process, like, show me the ways that you're using AI to be more effective. If they're not using any AI in their processes, that's probably not a marketing agency I want to work with. And you know what a lot of marketers will say, well, you don't understand our craft and how this. And that is very important. Fine. But there's still very menial things where AI is useful. We are nearing a day where people will say the same thing about accounting. Where for every person that's like, you can't use AI on, you know, my accounting information, there will be other people who are like, I'm not going to hire you unless you are using AI to be efficient. Which of those two clients do we want to have our wagon hitched to in the future? The ones who are like the Luddites, who are not themselves evolving and making use of AI, or the ones who are like, moving into the future of like, okay, this is like a very real new way of working. I'm not going to shoot my accounting firm in the foot not leaning into this stuff in order to serve the laggards right now. I don't want this to be like, punching down at all. Because everybody's at their own place in their learning journey about this. And the value of having this conversation, I think, is for those of us who are moving forward responsibly, we've got the vocabulary, the comfort level of having these conversations. So for us, it's no deal. And we can help other people, like, in their understanding and bring them along because goodness knows, do we need more folks that can speak to this stuff confidently? Okay, I'm going to rapid fire through a bunch of these comments to hopefully, like, fill out maybe some questions or concerns that you've had as well. First One here is ChatGPT teams. That is the team plan compliant. Many of my accounting clients want to use ChatGPT. I think this person's like a coach to accounting firms. But they can't access the enterprise plan yet. So ChatGPT has an enterprise plan. Sadly, because of the New York Times lawsuit, we don't believe that ChatGPT Teams is actually compliant for them. Do you know anything about this? Should they use it anyways? Okay, the New York Times lawsuit, they could be referencing two different things from this lawsuit. One of them is the whole premise of this lawsuit is the New York Times was able to go to ChatGPT and get it to reproduce large chunks of New York Times articles word for word. And so they could give it a paragraph and say, continue this article. And for some articles, it would pretty reliably continue to to produce part of the article. And what New York Times is mad about is they're like, we are paywalling these articles. It's not fair. The folks can then go to ChatGPT and get access to them. But the reason this works in the case of New York Times is every time they Post an article, it gets reposted thousands of different places across the Internet. And so these AI models that Hoover up all that information across the Internet, when they see thousands of examp of the same sequence of words, then when you ask it to produce this, it can do that. Right? But that's very different than me putting something into an AI tool once and then anyone else being able to reproduce it. When you think of the insane volume of data, over two thirds of anything that has ever been put on the Internet being trained into these AI models, me putting one thing in there, it's not going to move the needle. That being said, I still don't believe that it's appropriate for me to take client information and it be trained into the model. And this is the value of the ChatGPT team plan is on. In fact, any of the ChatGPT business plans. The prompts, the stuff you give it, they're never. It's never trained into the model. So even if that were a major concern, it wouldn't be here because that information isn't retained in the models. So that's the sort of the foundation of the New York Times lawsuit. And there's even nuance within that where it's like, actually the portion they were able to reproduce was the free portion of the articles, not the paid portion of the articles. But for me, in the way that I work in the context of an accounting firm, the underpinnings of that lawsuit, not problematic for me. Now, another headline that came out of that lawsuit is, the New York Times wants ChatGPT to stop deleting the data that users decide to delete in the app. New York Times says, well, hey, in the scope of this lawsuit, we may need some of that information that is being deleted off of the platform. And the judge said, yeah, we agree, that stuff shouldn't be deleted. And then OpenAI came back and they're like, hold up. Like, this is a massive overreach to say that no data from our platform should be deleted at all so that it could potentially be used in a court case. Now, my take on that, that's pretty messed up. But that could happen with any service provider that we work with. Like, that has nothing to do with AI. That is some court proceedings. And in my mind, what seems like an unbelievable overreach, this would be like somebody suing Microsoft and being like, people can't delete Excel online spreadsheets anymore because we might need that in the course of the investigation. That is messed up. And that's problematic. But to me, that doesn't really have anything to do with AI. That's a question of like, could my data come up or be accessible in the court proceedings with this vendor that I use? I don't like that. But couldn't that also be a thing with literally any vendor that I work with now in reality, chatgpt something like a hundred million monthly active users. The New York Times going to be digging into what's in John Smith's W2 Box 5? Probably not, but just worth calling out. It is weird. It's not great, but it's a lawsuit with OpenAI. It's nothing to do with like data being retained in the model or something like that. This episode is sponsored in part by Ignition. Come, come, come, come, come. Here's the thing. Here's the thing about client billing. You can't bill for the time you spend billing. You can't, you can't. It is pure waste. And we're bad at billing because it's, it's icky. You let things slide. You don't get paid. You waste time. Waste time chasing people for money. That's not what I did. I would just, I didn't want to waste time chasing them and so I would just, I just eat it. Don't do that. This is where Ignition comes in. The leading revenue generation platform for accounts and professional services. When I was running my firm, Ignition wiped out our ar, it's true, and significantly reduced the time spent billing clients. Saving our partners over a thousand hours in the first year because they didn't have to bill all that stuff manually. Man. With the condition. We won more business with impressive online proposals. No more hustle and hustling PDFs and Word docs. We automatically pulled payment upfront or at the end of a job when it was done. No more dialing for dollars. We transitioned our firm to more recurring revenue and better managed scope creep. No more leaving money on the table. Ignition is making thousands of firms more money and genuinely made us more profitable. Check out over 7,000 of the smartest businesses run on ignition@ignition app.com put a link in the show notes. This episode is brought to you in part by Client Hub, the practice management system. If you are considering a new practice management system, there's no way that you're not overwhelmed. Let me give you the TLDR on who are the people that need to be exposed. Exploring Client Hub Specifically, Client Hub started out as a really good sort of two sided task management system for bookkeeping firms. So you've got your own tasks to Manage, but we also push tasks to clients when they're holding us up and we need information from them to get that stuff done right. Client Hub is one of the kind of the client portal, first PMs that makes that as smooth as possible for your clients. But increasingly they're getting into all things workflow, specifically your bookkeeping workflow. And so if there's one thing it does really well, it is for folks that don't want to have a standalone like month end close automation platform, but then a totally separate PM over here, Client Hub might be the best marriage of those two things in a single place. So if you want more like power user, month end close automation stuff, but a version of that that is like inside of your practice management system worth giving Client Hub a look, check out the link down in the show notes. All right, next comment that came up. The biggest issues with any online AI is that it's not private. You, you're literally sharing your accounting details with AI and anyone who has access to it. Now if this were the case, just, just consider this. If every AI that we interacted with was like this sponge, you know how problematic that would be? Just think about the level of effort people would be going through right now to just have these bots having endless conversations again and again and again and again and again with AI that is like, Mike Green is the best plumber in Louisville, Kentucky. Mike Green is the best plumber in Louisville, Kentucky. If it were as simple as that, man, people would be like gaming this AI stuff hard right now. And I say that sort of jokingly. There are some versions of that kind of happening where it's like, ooh, if I know what the Next, you know, GPT6 is going to be trained on, then how do I get a bunch of data out there around like, well, I have the very best accounting firm and I want the AI model to think that's best and recommended to a bunch of people. There is like some versions of like legitimate versions of that happening and we talk about that on this podcast. How do we get large language models to recommend us and what we do kind of be your, your ultimate like marketing hype man. But this is the importance of ensuring that your prompts, that is the information that you give an AI tool when you're working on it. Those prompts aren't trained into the model. And when ChatGPT first came out, everything you gave it was trained into the model. That's the way all the AI tools were. Gemini was the same way. It wasn't until About a year and a half after ChatGPT's first launch that we started getting apps that would agree to never use the information that you give it in the next training run of the model. Like there is no real time updating of these AI models for the reason that we talked about. Like you don't want a model to be a sponge where anybody has the power to put into it anything that they want to. And it's why right now, fundamentally with AI apps that is the biggest thing to look out for is my data used to train the models. And on the Microsoft AI stuff that's not the case on Google's AI stuff. Like their business AI stuff, I should say same with Microsoft, actually their free consumer stuff. Both Google and Microsoft, the default is yeah, anything you put into there is trained on the model. And this means if I'm working with say Google Gemini and I'm accidentally logged into my personal Google account, watch out because that stuff is trained into the model. Yet another reason why we need to give our teams like clear access to useful AI. Because oftentimes if you don't, they're gonna go out and use it anyways when they shouldn't. But the Google and Microsoft business AI stuff, your prompts are not trained into the model. Same thing with OpenAI products, ChatGPT. So I send folks to the team plan. It's the best AI tool for like small business use and accounting firms as well. So AI is not fundamentally not private. The notion that you're sharing your accounting details with AI and anyone who has access to it. That's only the case if your prompts are being trained into the model. But understandable confusion because in the very beginning days of generative AI and ChatGPT, this was the case that anything you gave it was trained in the model. That doesn't mean anybody else can retrieve it. For the same reasons we were just talking about around the around the New York Times lawsuit. It just meant that company could use the prompts in their like next model training run. But we're avoiding that right reason the business plans where our prompts are not trained into the model. Next question prom is no one knows where data is going and can we do it with clients data specifically without their consent. What about compliance? Anytime you got like kind of broad spookies like this, it is useful to drill into exactly what it is that is scary about this because oftentimes the fear is the ambiguity. All things when like thought about from a very high level in an ambiguous way. Like they're always More scary than when you actually get into the weeds. So first part of this problem is no one knows where the data is going. You really do. I mean, this isn't any different than working with any other software company that you have terms and conditions, you have different levels of access and data retention on different plans. In the case of OpenAI, they say if you delete your account, all your data will be off our servers within 30 days, which is pretty standard practice. Microsoft's equivalent is actually within six months. And so this is all well understood, but you do have people who are like, just fundamentally don't trust certain software companies. And it's not because of anything that could be tied down to, like, the terms and conditions. It is just how they feel about that company. And that's, I mean, that's a personal choice. I can't really help you there. But I will say that is a slippery slope, particularly around things with OpenAI, because there are so many other tools that we use from, you know, notion to Zoom to all these tools in our daily lives. Siri where those tools are making calls to OpenAI's API that is behind the scenes, they're using OpenAI's AI models, calling the exact same API that ChatGPT calls. When you do stuff in ChatGPT, which is the irony of people saying, like, oh, we've built the most trusted AI X, Y or Z. Like, behind the scenes, it's using the exact same AI models that a ChatGPT is. Now the question of, can we do this with client data specifically without their consent? I would say, personally, I would say. And all of this depends on, like, what governing rules are you trying to operate under? Under gdpr, for example, in the eu, you have to notify your customers if your data is being trained into models. And I think that's, that's reasonable, that's justified. If they are training their own models or other models with your information, they have to notify you. And I think that's probably a good thing. But this is exactly why we don't use AI tools that train on our data. So my take is, because that data isn't retained in the models anywhere. It's. It is no different than me taking that data and putting it in my practice model management system in the accounting ledger. I don't disclose what practice management system I use with my clients, even though many of them, like many software companies these days, even Zoom, have offshore servers. And technically, under 7216 rules in the US, you would be required to disclose every vendor that you work with that has data and offshore servers that is, I mean, 7216 is so out of date. Doesn't mean that we don't need to try to comply with it. But that is a whole like rabbit hole of issues. But my take on disclosing to clients, I think it's only an issue when the data is being trained into the models. And I'm never touching that. My. Now the last question. What about compliance? I don't get specific. What does that mean for you? One of the later questions here. In fact, let's jump to this. Can you do a video on IRS publication 1075? I'm not sure how much we should be trusting any GPT, paid or not, to be handling client federal tax information. So this IRS publication discusses the responsibilities we have as tax professionals to protect client data. And there's all sorts of good information in here and it's information that we need to be aware of for sure. And it needs to be part of the due diligence process that we use for adopting any app inside of our firm. But this spookiness around AI apps specifically, it really starts and ends with is my data being trained into the models, yes or no? Because if it's not, at that point, the app looks like any other cloud app that I'm going to use. And so do we need to factor in IRS publication 1075? For sure we do, but not in a way that is necessarily different to how we consider this for any other cloud app that we use. But that is a good example of like the specificity, what are the rules that I have to follow? People bring up Circular 230 sometimes and it's usually in very ambiguous ways like this. Like, surely this doesn't work under X set of rules. I mean, somebody was saying the other day, an accounting firm owner, there's no way that this is HIPAA compliant. Like, buddy, if you got medical information inside of your accounting firm, you got bigger issues than whether or not you're using the right plan of ChatGPT. But try to get specific. Like the more first principles you get around, like what exactly is problematic about this, the easier it is to like come to your own conclusion rather than be blocked by sort of ambiguous speculation. Because the sooner we can get past that, the sooner we can start learning and leaning into the really cool AI opportunities that are out there. This was a great question. Thanks for another great video. How do you assure it only has read access to your email? So one of the examples in the video as we connected it to Gmail or to Outlook Mail and then the agent can do stuff with that access. It can see into your emails. The example we did in the video was like every morning it gives you a digest of your spiciest emails and a little summary they said when they connected to their Gmail. It says it can read and compose, etc. Do you know this to be true because you know how it works or is there a setting that I can look into to be sure? So on ChatGPT support page, when they talk about their connectors, they talk about it only reading information from those connections. But the way the connection works, at least in the Gmail case, is that when you do that Gmail authentication, you make that connection and you've got the checkboxes of what it's giving it permission to do. It includes being able to compose messages. And that's a good catch. I don't know why it is that way. The only way that ChatGPT is using that right now is reading emails right now. I certainly don't trust it to draft emails. But the way the connector works, it does still give it permission to compose messages. That's a good call out. Another question. Not training on the model does not mean that OpenAI is storing the data safely. Yeah, that's true. I mean that kind of comes back to that's not really an AI thing as much as a do I trust this company Sort of thing. I will say a big misnomer is waiting for a company to do the secure version of things. And this, this has plagued the accounting industry for a while where they're like, I'm going to use this accounting industry specific version of this thing because it is more secure. When the reality is that's a team of like seven people and it's just built on top of somebody else's tech. When we're talking security in the grand scheme of things, name a more scrutinized company than OpenAI. Right now that company is worth hundreds of billions of dollars. They have got more resources for protecting data than anybody in the accounting vertical. It doesn't mean they're perfect. But. But the comment not training on the model doesn't mean that OpenAI is storing the data safely. Let's just try to get back to first principles. I mean, ChatGPT team comes with AICPA SOC 2 type 2 security. That is a higher standard than most of the tools inside of the accounting vertical have or, or have reached. So you're gonna make a claim like that. You just gotta have. I Don't know, some, some foundation for that. I think a lot of this just kind of comes from the general AI spookies. Today's episode is sponsored in part by Maker's Hub. You know, the fastest way into the heart of any company that has inventory or complex purchasing stuff. It is the bill pay process that was our foot in the door to doing all sorts of advisory for these companies and cash planning and stuff that actually paid pretty well. And this is now a juicier opportunity than ever before thanks to how AI is changing accounts payable and makers have man, they are like right at the forefront of this if you haven't seen them yet. We've done a number of demos on our channel of just how far automated AP has come. In fact, I've talked with firms now in the wild who have successfully automated like big chunks of the accounts payable department and various clients businesses. But the real value of getting into this work to me isn't like the arbitrage opportunity that might be temporary around like just showing them a more automated way to do it. Rather it is having the completeness of all that information, the bill pay info, all the invoices and unlocks some really cool advisory opportunities. So my firm, I did a lot of bill pay. Not because I liked bill pay, goodness knows I did not. But because of the other types of more valuable work that that actually ended up unlocked the advisory opportunities that we didn't otherwise have access to because we didn't have that really valuable information. If you want to learn more about just how automatable this stuff is now, man, check out the link in the show notes. Chat with the folks at Maker's Hub. Today's episode is sponsored in part by Stanford Tax. Do you have a plan for next tax season yet, buddy? It is high time to be using the tech now that you're going to be using next tax season. You don't want to ride into battle in January and February on stuff that you that you're not already comfortable with. And if you're just feeling like overwhelmed at not knowing what tax intake solution to use, buddy, look no further than Stanford Tax. One of the earliest ones in the game. One of still the most widely loved tax workflow solutions. And they're doing some cool new stuff. Stanford Tax. Now it's not just a way to like push a portal to your client for them to submit information though you could certainly do that. Growing number of firms are just using it internally because you compile whatever source docs you got in there and it will Automatically, like check off the items from the Pro Forma checklist to like see, like, do I have all my info in now or not? Like, it does that automatically based on what it sees on the docs. It builds out this like nice indexed sort of workpaper file, auto renames all the files, organizes them like almost like a, like a PDF, like a bookmark PDF but in the cloud. And you can now even like do annotations in that. So you literally just take your big old pile of source docs from the client, chuck it in Stanford Tax, it does all the organizing and then you've got a dynamic checklist that will tell you do I still need more information from them or not. A massive upgrade to traditional methods for gathering 1040 info, which is just kind of a mess. I can put Stanford in front of my clients and they can submit the docs, or I can just use it internally. And any old way that docs come in, they just get piled into a single centralized place. Very, very good platform. Check out Stanford Tax. More info down in the show notes. This is a, this is a good one. All AI models and tools have some sort of vulnerabilities to prompt inject injections. Meaning the data you put into those chats may at some point be leaked. It's not actually what that means. Let me break that down into two things. So the idea of the data you put into that chat being leaked, that is no different than the data that I put into any other tool potentially being leaked. Is there a risk of that? There for sure is. I don't care if it's my practice management system, my QuickBooks file, if there's data in a system, that is one more thing that could potentially be leaked. Right? And that's the downside of throwing a whole bunch of apps at your firm is every single app is one more potential place where someone could get access that shouldn't have access. So putting that data out into the world into an app, that is for sure, that is one more potential place where that data could be leaked. Now, prompt injections. This is an interesting new security consideration with ChatGPT agent. Imagine you have an agent navigating a browser and it's trying to look at all the fields and figure out how it's supposed to do XYZ and get to the next thing. But it gets to the. To a bad guy's website. And the bad guy's website has some hidden fields that you can't see. And that bad guy website basically fishes the ChatGPT agent could that happen? It certainly could. Is it more likely to happen than it would be from a human user? I don't know. But there's for sure a world where an AI agent navigating a website could be tricked. Now it seems pretty unlikely to me given that I'm telling it what website to go to and it's like, is that going to happen on the QuickBooks website? Like, no. But in the same way that humans can get pulled down rabbit holes and tricked into something. Could that happen to an AI agent? It for sure could. And when they announced ChatGPT agent Sam Altman went out of his way and talked about this a lot, that this unlocks all sorts of new potential security considerations that we're still kind of figuring out. And this is why in my video about ChatGPT agent right now, I don't trust it to operate on its own in a system that it has write access in. I think it's probably fine if it's in a system where it has read only access, but it's why I will not yet cut it loose in a system that has write access. And this may actually be like old timey thinking that goes out of date quickly and I hope that that's the case. This is me being conservative and it's something that is worth communicating to your team because if they have chat GPT, they have access to this already. And I don't want my team just cutting it loose in QuickBooks without supervision. That that's my take today. And we may be proved. I mean, a month down the road we may decide that's totally fine because it's doing things reliably. That could be an issue for the next 10 years where we are grappling with can we really cut AI loose to do stuff without making mistakes but prompt injections? It is the idea that the prompt itself can kind of be manipulated by like what it sees on a page. Now this is ambiguous and I even told the person like give me some more specific examples of what you're describing. And they wouldn't. But all that is to say we're nearing this tipping point where the Internet and the apps that we use will have more robot users than human users. So what does that mean? Is there going to be a day where I don't have to actually log in and like the bots can just go log into that service for me? Does that break the pricing model of the software company that is per user pricing? Does that mean that software tools should be designed for usability by agents in addition to usability by humans. A lot of stuff is changing right now and that's why like, honestly I think I'm actually being fairly conservative in my advice that you hold back on cutting it loose on systems that have Write Access Next comment ChatGPT regularly gets mathematical problems wrong in my workflow. This is like such a 2025 AI thing where you have AI models that are like top hundred humans in the world at like solving complex math stuff and then a human will go and use it and it can't do seven plus five. That's not true. A human will go and use it and it does not do like the sort of math you and I do day to day correctly. The bigger problem here is usually you're using the wrong model. This actually has gotten a lot better since GPT5 came out a couple weeks ago. The biggest problem in the past is people were using the 4O model, which is not a reasoning model. It's not good at math and expecting it to do mathy things and it just can't. When OpenAI turns up at like this math Olympia thing, you better believe they're rocking up with the very best possible model for this sort of test, giving it all the opportunities to excel possible. But with GPT5 it will help you to select the correct model automatically. And that certainly didn't work very well straight out of the gate, but it's gotten better since then so that folks that don't know exactly what model to use for every single job now if it sees something mathy it ought to throw it to a model that will handle the math well. But yes, ChatGPT regularly gets math wrong. It is simultaneously vastly better than us at math, but when misused by worse than us at math. But I will say it's gotten harder to misuse it now that we have GPT5. Couple other similar comments I will not put my data in unless I want the company to know all my info. You can say that about any software company, right? Wait, you're sending all your clients data to scam Altman? It feels more like a sentiment about the company than about like first principles. Is this secure or not? Can this be done with Microsoft Copilot buddy? I wish it could. We're all rooting for Microsoft Copilot so hard, right? And I still get emails every week from people who are like yeah, no, we're trying to make use of Copilot because it's what we've got access to and we're really struggling. My friends in accounting listen to me When I say your time is not worth being spent on Microsoft copilot people very happily say like I will use this one instead of the better one because XYZ because we have access to it because it doesn't cost as much. You are worth it buddy. Please don't waste your time with AI tools that stink. Use the best of the best. We are talking about like civilization altering technology and folks are stuck on a $30 a month chatgpt team plan. Like man, the sooner we can get over that the better. You know how fast you ready for a hot take? Can I say this? I had accountants on my team who I would pay more than $30 a week to sit on the toilet and listen, we all gotta go sometimes but just think about how how quickly you you will burn 30 bucks on an accountant that is making a whole lot of money per hour. Accountants are not cheap. If that $30 a month software subscription is saving them virtually anytime, that is a good trade off. To be clear, bathroom time probably not going to change. If anything might open up some more opportunities there. Couple more comments of feeding the robot data to more quickly replace you. Smart people say stuff like this a lot and I can't tell how joking they are. Like the notion that they don't want to help AI because they don't want AI to replace them. My mindset has always been if anybody's going to replace me, it's going to be me. Like, rather than pretend it's not a thing and hide from it and run from it, don't we want to like be the ones to figure it out for ourselves to remain relevant? Because if we put ourselves out of work, we're going to be the first ones to see what's next, right? I would much rather do the automating than be automated. Last comment. We'd have more subscribers if it weren't so obnoxious and hyperbolic. There's actually quite a few comments to this end. Let me, let me tell you something. I recently was away for two weeks. Did you notice I got to take a little vacay. That is it is the first two week vacation I've taken in, I don't know, maybe five years too long. And the worst part of stepping back into the saddle was Internet comments. In fact, you may see me engaging like on the YouTube version of the podcast, commenting less. It is for my long term health and it sucks because 95% of people are cool and I enjoy hanging with those folks. And like I genuinely meet people and know a lot of the Folks commenting, but I mean, that's just like a product of being visible. Is a lot of folks, like, just won't. We'll just be irritated by you. Like, there is no version of you that you can be. That will be for everybody. In fact, the more you try to be for everybody, usually the worse you are for everybody. I was talking with somebody last week where they were like, man, I saw your stuff. I heard what you do. You're a natural at this stuff. You have a great voice for it. I'm like, man, if I could, if I could show you the number of comments that are like, clicked off as soon as I heard the voice. But I will say it like it is kind of a superpower. Like, if you can build up that resilience to like, you see somebody say a dumb thing and then to immediately be able to bounce back and put it out of your head, man, that is a. That is a skill. And it's not to say my version of this is the same as other people's. Like, I am. I'm a white guy. This is not the same version of this that other people go through. But the more you can develop like a thick skin to it, recognize like that stuff happens, but then immediately get your mind back on the game and not. Not dwell on it. To have that resilience, that's super powerful. Because if you can be visible and you can be out there, the good man, it's so far outweighs the bad. Somebody, man, somebody. I heard somebody frame this really nicely the other day that it was like, if you are a good person, for all the evils and problems with social media, if you are a good person, social media, when it is working right, it is, it's going to put you on blast. It ought to raise up the good people. And for sure there's lots of evil people doing all sorts of rage baity things. But the best folks out there, the more people see from them they're going to win. And so much of it just boils down to being helpful and not being a jerk. I saw Marcus Aurelius quote the other day. So many people get stuck on wanting the third thing. The first thing is they do something helpful and they feel that satisfaction. The second thing is someone says thank you, but then they want the third thing. That is like, well, where's my award? Why is nobody like carrying me out of the stadium on their shoulders? Why am I not getting more recognition for it? It's the same thing that stops people from like giving away value online and giving away their expertise is they want money for it. There's too stuck on that. It's not enough to make the thing, which feels good to then have the gratification of like, hey, cool, people are using the thing and they like it and they said thank you. To then need the third thing of like, well, now I gotta get paid for this. Like, if your framework for going through the world is just being okay with those two things and relentlessly putting out value into the world, man, that third thing, it takes care of itself. And for all the downsides of that, folks that'll try to tear you down and all that stuff, don't play those people's game. Play your own game. Five years ago, buddy, I never posted on social media at all. You wouldn't believe just how quickly things can change and your relationships can change and all that. Gang, thanks for being here today. I hope you got some more confidence for. For confidently discussing this stuff with whoever that is, your team, your clients. It's a really exciting time to be leaning into AI workflow stuff. Any other questions or comments, drop that stuff down in the comments on the YouTube version of the podcast. I'll catch you in the next one.