478 An AI Security Primer for Accounting Firms

AI buddy, it is happening. 2025 is the first year where it's like, I honestly feel that you are losing out on like big time opportunities for productivity gains if you aren't leaning into AI. Obviously there's lots of different types of AI. I'm talking AI assistants like the chatgpts of the world. And it makes the security conversation more important than ever before. What's the responsible way to use this? Can we use it with sensitive client information? Spoiler alert, alert. Yes. But we're going to dig into the nuance. The fact that there's still insecure or non secure ways to use AI doesn't mean that there aren't completely secure ways to use it. And I want to just organize all this stuff into a single episode so we can kind of COVID from soup to nuts a lot of the headlines and the information from the early days of generative AI and why those concerns were very valid and where they still apply and where they no longer apply so that you can confidently move forward into the future. Take my hand, let's do it together. AI is just a confusing thing and a lot of it stems from sort of the original sin that is generative AI. Like all forms of generative AI are trained on stuff that, that humans made that almost without exception. Humans never gave these companies permission to use in their trading data. The only exceptions to that is now you have some like, you know, certified organic sort of models, namely from Adobe where they were said to only be trained on, on assets and art where they actually had permission. But that is the exception to the rule. By and large, companies have just been gobbling up anything that they can grab in the sort of land rush that has been the years of just making the biggest AI models possible. And we're actually at sort of the end of that era where bigger was always better. They're learning now that models today are largely trained on like everything they can get access to on the Internet. And they've, they've actually run out of training data. And so now companies are getting creative about like, oh, how can we come up with more training data here or there? They're doing a lot of things with like synthetic training data, like take a really, really good AI model, create a bunch of fake data that was generated by the A, and then if we train that back into the model, like, is there a way to get even better results? Because the day that an AI can make something more interesting and more compelling than a human, then isn't that better training data than the meat bags are giving it this whole like, I mean, the fact that there is like some evil baked into generative AI, I think this is where a lot of the discomfort comes from. And this, the salty YouTube comments about, you know, OpenAI just being a data harvesting collection company, which honestly I don't know that there's many Internet, like big Internet companies that people don't say that about, from, you know, Google to Facebook. And they're not wrong. But ultimately, within the scope of using AI inside of your accounting firm, it's got to boil down to is it okay to use this or not, and is it only in certain situations, or can I use it in anything? And I'll just like, spoiler alert right now, tell you where I, where I am on this and then kind of give you the background on how we got there today. There's completely secure ways to use AI. In fact, when we're talking just about AI chat assistance, the only thing that makes an AI tool like that any different than every other cloud app that you use in your accounting firm is whether or not the prompts or the information that you give the app is ultimately trained into the model. That's the only thing that makes it different from every other cloud app that you use. Now, not all cloud apps are secure. So for any cloud app that we use inside of our accounting firm, or any desktop app for that matter, we need to have a due diligence process, taking a bunch of boxes to say here's kind of all, all the security requirements, stuff like, can we tell them we no longer want to use their service and they'll delete all of our data. Can we enforce two factor authentication, stuff like that. Like, we have to have a process for vetting apps that something more robust than, well, I heard he's using it. And the only difference with AI apps to how you vet every single other app is you just need to make sure that your prompts aren't trained into the model. So if I ask it a question or if I upload a client's financial statement to it or something like that, if prompts aren't trained into the model, then the considerations are the same as they are for any other cloud app. Now, if we go back to early days, Generative AI, every single tool trained the model on the prompts. So initially, like Chat GPT in the early days, there wasn't a way to use ChatGPT without all of your prompts being trained into the model. In fact, it wasn't until we got the Chat GPT team plan, which is still the one that I recommend accounting firms use. It wasn't until we got this plan in January of 2024, so we are well over a year now that we've been kind of talking about this. But it wasn't until we got this team plan that we really had a model that. Not a model, but a plan that was fundamentally secure for business use. Before that, that was where he saw people splitting hairs around things like, well, you can use it with non sensitive information, but you can't use it for sensitive information. And that was where the whole like AI usage policy and all that stuff came out. The notion that there's okay ways to use it and bad ways to use it. And that's like kind of flawed. Like I don't want to rely on a member of my team remembering to redact a thing or rely on them understanding that it's okay to put this stuff in, but not okay to put that stuff in. I just want the app to be secure for any use. And that's, that's where we're at today, the Chat GPT team plan. And we'll get more into this, but under no situation in the team plan are your prompts trained into the model, which means your due diligence checklist for using that app. It looks the exact same as any other cloud app. There is no AI boogeyman or anything really unique to it. Now let's get back to kind of the background and where a lot of the spookiness and a lot of the misunderstandings sort of come from. Because if we go back to the idea that, well, it's, it's trained on everything. I've had people tell me things like I'm not going to use ChatGPT to help me write my blog posts because I don't want my blog post being trained into the model. I don't want somebody else being able to access it from ChatGPT or something like that. And the reality of that statement is it's actually backwards is when you post your blog on the Internet, it is going to be trained into a model. If you use, you know, the Chat GPT team plan where your conversations are not trained into the model, then as you're working with ChatGPT to build a blog post, that conversation wouldn't be trained into the model. But if you ultimately posted it in a public place, in all likelihood it would be trained into the model. There is, there are like instructions that you can actually put on your website under like the robot text sort of parameters similar to like how you Signal to Google. Do I want this to be indexed in Google or not? There are parameters that OpenAI say that you can use to keep your website from being trained into the model, but there's a lot of debate about whether they're actually respecting that. And then also like every other AI company in the world, like in all likelihood they aren't. But the reality is in 2025, I mean, in most cases, and I'm just talking about knowledge work because like creative work and I'm thinking like more visual, like images and videos, stuff like that, that's a whole different ball of wax we're not going to touch on. But if I'm running an accounting firm, I want a large language model to know about my accounting firm. People find my stuff from ChatGPT, from Grok, from AI tools that they're working with. In fact, the way these large language models work is they basically generate a token at a time. And a token can be like a short word or part of a longer word. Basically all it does, and this is kind of the amazing thing about how these work, basically all it does is it looks at all the text before it and generates the next token at a time, probabilistically. So it looks at all of the training data in the entire model, and based on everything that it's ever seen, it generates the next little chunk according to the text that's been generated so far. So if it's a chat conversation and I ask it a question, then it's generating the next token based on the question that I asked. If it is midstream, like through a statement, then it's generating the next token of text based on everything above that. And it's purely done based on probability from everything else that it's ever seen. Which is what makes this such unbelievably like novel text, novel tech, because it is in many ways such a brute force thing, right? Like when GPT4 first came out, like, I still remember that first day and it was March of 2023, March 15th actually, and I got early access and this is the day. Maybe you haven't heard this story, but OpenAI did a live stream announcing GPT4, which was the first new model after ChatGPT first came out. And everybody's spooked. Like ChatGPT originally, they're like, this is so novel, so powerful. What's the next AI model going to be able to do? Is this going to disrupt entire markets? Well, they come out, they do this really impressive demo, and the last thing they do in the demo like this sort of a one more thing that we're kind of going to do for fun and is they, they quote, unquote, prepare someone's taxes. And I'm sitting there and at the time I still run a 40 person tax firm and I do like a spit take. Like, what, Hang on, what? Now, earlier that morning when I first got access. Now this was, this was two months after I sold my firm. I didn't have my firm at this point, but earlier that morning when I'd gotten access, it's March 15th, so I'm like, I made up this fake sort of client situation to run an extension estimate for them. And I gave that to the new model just to kind of like do a vibe check and see like how much of this can it actually do. And it built out this very impressive response that was like, oh, breaking down this, that and the other thing. And it was, I mean, it wasn't something crazy. It was not something crazy complicated. It was like a W2 earner. I gave the withholdings, number of dependents, the tax year, all that stuff. And I went to Twitter and I said, can somebody just drop this into their tax software real quick and double check what the actual correct answer is? And it got it correct to the dollar. It was perfect. Now that was just a text response. It didn't complete the forms, it didn't do any of that stuff. But just think about that. This is a piece of technology that generates one token at a time according to all the stuff that came before it, according to all the, the information it was trained on just on probability, it generates the next token at a time. And that somehow is like some version of quasi intelligence that can perfectly calculate an extension estimate in this specific situation. Isn't that wild? Like, when they were building these transformer models, nobody expected this to somehow be like smart. And since then, it's just continued to get better. But because it's kind of this spooky black box thing, it understandably makes people nervous. And there's this concept called capability overhang. That is the fact that a language model's capable of things that we haven't even really yet discovered. Until you're like, get into it and you give it sort of a certain combination of information, some prompts, and you're like, oh, we actually just found that it's really, really good at this specific thing. If you're an accountant, you know, a lot of things that AI still isn't good at. Right? Like early days. ChatGPT. It was very, very bad at math these days AI is like 99.999 percentile and ability to solve like competitive math problems. But what, what AI is good at and bad at is like not always super clear. It's, they talk about it as kind of like this jagged edge of, of what it's capable of. So like generating code, writing code is very, very good at that. Better than anything else arguably. And that's why it's having the biggest impact probably on like software developers. But this capability overhang the notion that we don't even know what it's good at yet creates this sort of mystique around AI that suggests, well, maybe we don't understand how it's actually using our data then either. But that's not the case. We do. Capability overhang the notion that we're still discovering what AI models can do is very different than not understanding how it uses our data. How it uses our data is clear as day. The fact that your colleague finds a really cool new use case, that doesn't mean there's something going on with your data being trained into the model or something like that. Those are two entirely separate things. A model is trained on a whole bunch of data and then at that point it's kind of sealed. There's other things you can do like, like fine tuning and other ways that models can be sort of quasi like trained or modified a little bit. But as you're interacting with it, you send it stuff, it bounces off a response and there is no stickiness to the stuff that you sent into it. Now AI companies or even just like the free chat GPT plan which is trained on prompts. So if your mom goes up and sets up ChatGPT on her phone and she's having interactions with it, all that stuff is trained into the model, but not like live, not as it happens, that data is rolled up along with all the other data and included in like future model training runs. So this excitement over like this sort of mysterious black box, boxness of it all shouldn't be confused with well, what about our data? How is it using our data that we totally understand? So some examples, some headlines that were sort of spooky that people still have in their heads. Back in April 2023, there was a pretty high profile case of Samsung software engineers putting their whole code base, their proprietary code base into ChatGPT. And at the time everybody's like, wow, this is like company secrets, like trade secrets are just being handed over to OpenAI. This isn't a tool for work and that was a dumb thing to do. It was because in that context OpenAI had permission to use those prompts for whatever they wanted. Now these days, how things have come full circle. If we look at, you know, the ChatGPT team plan where your prompts are not trained into the model, that same use case where a Samsung engineer is using AI to help it write some code, that ought to be totally fine. In fact, developers who use OpenAI's AI models in their software and so that developers are not going to ChatGPT, they have sort of a different end around and that is using the API. So for example, Notion. Let's talk about Notion. Any software app that's using OpenAI's models, which is almost everyone on the back end, they make an API call to OpenAI's servers and it looks just like interacting with ChatGPT. It is, here's the message and any attachments and the API sends back a response. It's the exact same API that sits behind ChatGPT. Whenever a tool is using the API, any data sent over via the API is not retained, is not trained into the models. And so that same company, Samsung, who were like why? This isn't a tool for work and they had this situation where their code was handed over to OpenAI. All of those same companies now partner with OpenAI and send data to their API all the time. The difference being data sent to the API, it's not trained into the model. Just like using ChatGPT team plan, your prompts are not trained into the model. But again, like a lot of the headlines here at the time that was a very legitimate concern. Today's episode is sponsored in part by ClientHub, the practice management system that delivers wow seasons year round, not busy seasons. Seasons that will dazzle your clients for two reasons. One, Client Hub is drop dead simple man. If you're too tired of a practice management system that just feels like you gotta have a of piece PhD and learning something new. Client Hub, you hop in, the vibes are great, it just all makes sense. But two, remember, there's a second reason Client Hub offers a game changing client experience. That is you can actually get clients to like collaborate with you and be helpful on the platform. Cause the portal wasn't this like afterthought? Bolt on now. Very cool. Recent development in ClientHub Magic Books review Twinkle sounds so ClientHub syncs in the transactions from your QuickBooks file. You can run an AI analysis on it like from inside of Client Hub. It'll highlight any potential issues you can Mark like on the balance sheet in the P and L, you mark each account one by one, whether or not it's been reviewed. But if you want to troubleshoot, it'll pull you straight back into QuickBooks like so you don't have to learn like a whole new troubleshooting tool. If you'd rather sort that stuff out in QuickBooks Other cool stuff it'll identify automatically expense inconsistencies, purchases without a payee purchases and parent accounts. Why is why can we not disable that QuickBooks Large transactions, new accounts, new vendors, new customers. It's magic Books review inside of Client Hub I think you might like it. Learn more about Client Hub Check out the link down in the show notes. This episode is sponsored in part by Verito. Here's a hard truth incoming Oof. Bomb. If you're still managing your own IT relying on generic IT guy, you could be non compliant with FTC safeguard rules, IRS publishing 4557 and WISP requirements. That's exposure you don't need. But enter Barado's barrettguard, Barrett Guard. Can I get some feedback on that? No? Okay. Barrett Guard gives firms fully managed IT that's purpose built for compliance. That means your systems are monitored, patched, protected and documented continuously. Over 1400 accounting firms rely on Verito to stay audit ready. Gang, the reason we're actually talking Verido here. Can I go off script? About six months ago, I put out a post on social media and I said, oh, first party hosting from the usual characters you've heard of. Not great. What's everybody using? What do you like? And I saw a level of love and fandom from Verdot and I'm like, hang on, why is everybody so enthusiastic about like a hosting provider? I mean, don't get me wrong, this is important. And so Verdo, they have had a place in my recommendations ever since then. Because firms are just like, do it. You won't regret it. Whether it's I'll assert Drake QuickBooks Desktop. They speak your language because they work with a bunch of other accounting firms. Okay to schedule a free consultation just to learn more. What do you got to lose? Check out the link down in the show notes. There were some other news stories early on about people being able to retrieve medical images from ChatGPT. In fact, one person even came up with a medical image of their own and was able to get it from like this, this AI data set that was out there on the web. It wasn't directly from ChatGPT, but this kind of Spooked off and kicked off another flurry of like, oh my gosh, like anything sensitive that you ever put in there, it's just going to be there and other people are going to be able to retrieve it. Well, the story behind this was this person's medical images had been sold to a third party company. And this is actually very common practice for redacted medical data to be sold to Google, to all sorts of people. In all likelihood, UNI's medical data has been sold a bunch of times, but it's redacted under HIPAA rules, under GDPR rules. In general, sensitive information, once it's redacted, that is removed. Like anything identifying has been removed from it, it's generally no longer considered sensitive information. So your doctor. And I'm, I'm getting in way above my head right now, but this is why like medical companies and pharmaceutical companies can buy a whole bunch of medical data is because it's been redacted. So in this case, this person's medical images had been sold to another company. They used a tool that let them do an image search. So basically they could provide it with an image and then the service would look into the AI data set and find any matching images and they found it. And this data set was used to train ChatGPT. So you can see the connection there of like, all my medical images have been trained into ChatGPT. That's obviously a very problematic thing. But it was redacted and that happened because of, I would argue, other evils going on where your medical data can be sold to whoever wants to buy it. Now, probably the most commonly cited court case, which is still an open case between OpenAI and New York Times, this is the one I hear most common as sort of the argument for why you shouldn't be using ChatGPT or AI or something like that. New York Times is suing OpenAI because they provided some examples where you could prompt in ChatGPT the first part of a New York Times article and it would produce, I don't think in any case like the entirety of the rest of the article, but like the next couple sections of the article. And so New York Times is saying like, well, this stuff oftentimes is behind a paywall. We won't, don't want people to just be able to be like, hey, give me the rest of this article from the New York Times. And then it produces it. But if you remember how these models work, they're probabilistic. And so if I gave it two paragraphs of A New York Times article. How likely is it that those two paragraphs existed anywhere else on the Internet in a different context where the next paragraphs wouldn't be a continuation of the same article? Right now, devil's in the details here because every single time a New York Times article is published, it is reposted to thousands of different websites. Namely the free portion of is reposted all over the web. And so if you think about what ChatGPT is trained on and all these tools, if it sees 10,000 examples of this sequence of text, then yes, it's very likely that it will produce some more of that same sequence of text. But what it wasn't trained on by and large was the, like the full version of paid articles. Unless other people are posting it online then, then yes, to a smaller degree it would have scooped up that information as well. But it is, I mean it kind of calls comes back to sort of the original sin of AI. Is it okay for even just the free publicly available portion to be presented inside of ChatGPT? It's not too dissimilar to a lot of folks not liking Google. Like Google absolutely destroyed media companies when they just became a, a more efficient, faster way to find information. This is happening again now with Google's AI summaries where it'll answer questions for you without you even having to click into the search result. It's kind of messed up, but it's also like a toothpaste in the tube type of situation where you're like, I don't, how are you going to put, how are you going to put the AI genie back in the bottle right now? And that doesn't make it right. But I look at this through the lens of like, I'm a business owner and I kind of got to skate where the puck's headed. And this, I mean this becomes more of a personal decision. But to be a never air, that's not, that's not realistic. Now some other sort of vernacular that there's some confusion around the notion of open source models. Folks think that OpenAI creates open source models. They used to, they don't anymore. But an open source model doesn't mean that people have access to the training data. So the highest profile examples of open source models right now would probably be Meta's AI models. Meta's, the company behind Facebook. And then Deep Seq's model. DeepSeek is the Chinese company that released DeepSeek R1. And it was this big, like really powerful open source model. If it's open Source doesn't mean you have access to all the training data, but it does mean that you could take the model and run it on your own hardware. I say that, but the like, all of the models you interact with every day are vastly too big to be able to run on any hardware that you or I own. There are itty bitty little baby models that can be run on things down to a smartphone, but the reality is they suck. And all of the things that make today's most powerful models cool, these itty bitty baby models can't do. But people hear open source models. They hear, or they hear even just open models and they think of that like a box where everything that they ask the model is getting poured into the box. It's, that's, it's not like that. Open source refers to like usage rights and availability of the model. But this then ties into the discussion around local AI. Well, is local AI then the secure solution? Because everyone has this sort of, I don't know, distrust of AI companies. And I would tell you that it isn't for the same reason that today it's less secure to store your clients files on your own server than it is to store it on Microsoft or Google servers, because somebody can literally come into your office and haul off with that server. That's an easier door to kick in than Microsoft or Google's doors. And the same applies here for AI. And we're just cruising right past the whole question of like how much hardware would you actually have to have to have to be able to run powerful models? Like it would be astronomically expensive, but let's just throw all that out or let's say that we're going to get by with really crappy bad models that can be run locally. Is that really solving the security question or is it actually creating a bigger security risk? Because local network infrastructure generally stuff sitting on your own server I would argue is generally less secure than going to Microsoft or Google for the same thing. Now there's actually a lot of vendors preying on this right now where they're like, we figured out the problem. The problem is cloud AI is inherently insecure. And so we're building our own, our own models that are our own like custom trained models. And they are seizing on the fear of like, I mean there's, I mean every week now there's still a headline of well, somebody sent a bunch of medical images to the free chat GPT or to Grok or something like that, and it got trained into the model and it's like, yep, they did. And that was pretty dumb. But again, the fact there are non secure ways to use AI doesn't mean that there aren't totally secure ways that we can use it for business. But these companies will kind of prey on these headlines and reshare these headlines and say local AI is the solution. But what they're actually saying is it's their own local AI that they run on their own servers and you can have access to it. But now we are saying, no, I would rather use AI with this small little startup company that has three employees instead of going to Google or Microsoft or open AI for AI models like that. That three companies, that three employee company is not more secure than the big guys. And once you like reason through all of this stuff, it comes back to. In fact, there's a recent article that just said it. It was actually TechCrunch from just a couple months ago. You shouldn't be uploading your medical images to AI chatbots. And at one point they said in the article that you have to trust that we're forced to trust the company at their word. So if the AI company says like, no, we don't train the model on our prompts, there's a lot of people who are like this where they're like, well, I just don't believe you. And that's. I know that's a hard thing for me to reason around or help you to get over, but I would try to help you understand if you got a beef with, say, OpenAI right now and you don't trust them, and then you come over to your vendor over here and you're like, no, I want my vendors to implement AI for me to do X, Y and Z with. I'll let you know a secret. They're almost all using OpenAI, even Intuit. And so, like a pretty cool, novel thing that's actually happened in this era of generative AI is in the past. And I won't belabor this too much because I think we talked about it on the POD last week, but in the past, everybody had to train their own AI models. And it was astronomically expensive. And so only the big dogs could do it. They were the only ones that had all the data. And so people like you and I would never have access to cool AI models. Now, in this current era of generative AI, you have companies investing hundreds of millions of dollars to train something like GPT4O and Intuit's using it, but you and I can use it just the same way. That's so cool. That's such an opportunity. But there's still a lot of confusion around like, well, I don't trust OpenAI, so I'm going to have my other company do this. The reality is this other company, they don't have $100 million to train their own model. They don't have any of those capabilities. There's a reason why OpenAI has like created this entire category. If it was just a question of, of money, then others would have beat them to it. It's also a question of know how. And you better believe the people that know how to do this stuff right now are absolutely making bank. Mark Zuckerberg is trying to poach everybody at OpenAI right now for good reason. But that other random company that you work with, that's much smaller, they didn't figure out how to train their own models. In fact, I think it's really funny. You look at the bigger, more mature companies like Notion or even like Intuit, they're like, yeah, we partnered with OpenAI, we use all that stuff on the back end. And then you see other smaller companies who are like, no, no, no, no, no, no. We've trained our own models for this or that. And we're like, did you, did you really Tell me why Notion isn't doing that, but you are. And so even if you have fears and reservations and all that stuff about using ChatGPT, just understand. But behind ChatGPT is an API hosted by OpenAI. And it's the exact same API that all of your other tech is using to enable cool new stuff with AI. Today's episode is brought to you in part by Team up, who helps you hire direct in the Philippines. What is hiring direct? Hiring Direct means you own the relationship with that person. You're not paying a middleman, a middle person. You're just paying the person that works for you. And that's all they do is work for you. It's there, an extension of your culture. It's no different than hiring somebody in another state. You're responsible for training them. They're just part of the squad now. Things to figure out. The first time you're doing that, what does it look like? There are a number of things to work out. If you do tax work, you've got some disclosure requirements. But that's why pulling in someone like Team up, where this is all they do, is really helpful. In fact, Team Ups been around for 26 months. These are, these are real numbers. They've now placed 250 hires into firms 98 bookkeepers 98 bookkeeper slash accountants 80 senior bookkeeper slash accountants 42 tax accountants 11 team leads slash managers 10 EAs they recruited for 135 firms. Oh, and 184 of those hires are still with the firms. That's fascinating. 184 out of 250 hires. That's a pretty good number. And team ups even got like a six month guarantee where if that person doesn't work out for the role that you had defined, they'll find you another person. Team up. What they're really good at is direct hiring. If that's what you're looking to do, schedule a call with them at the link down in the show Notes this episode is brought to you in part by Financial Sense, the practice management system. If you've downloaded my app database, followed my app recommendations, Financial Sense has been one of my top recommended practice management systems for you. While now, Financial Sense gives your team one place to track client work, assign tasks, follow up with those clients, run reports and hit those ever so important deadlines. Thousands of firms use Financial Sense, including Fiscally professional Corporation, a 30 person accounting firm. Jessica Wong, a director of the firm, says Financial Cents gives us the pulse check we need, ensuring we stay on track. The checklist templates have standardized our onboarding process process very effectively. It's good and we've done a number of demos of Financial Sense on the YouTube channel. If you haven't seen it yet, right now, teams of two or more, you're going to get 10% off the team or scale annual plan. Take Financial Sense for a spin for free at the link down in the show Notes I don't want, I don't want to like diminish the concerns or anything like like they're completely valid because still there's a host of ways that we can use AI that is not appropriate. Namely on like free plans and even even OpenAI's plus plan which is $20 a month by default. It still trains your prompts into the model. But there are completely secure secure ways that we can use this stuff. Now Microsoft and Google have their own sort of E AI tooling within their business products and that stuff is all every bit as secure as anything else that you use inside your Google workspace or your Microsoft 365. Problem on the Microsoft side is it just sucks. It's so bad. Copilot's useless. The Google stuff has gotten really good lately. Probably I don't maybe as good as ChatGPT Chat GPT just as an app does a lot of really, really compelling things that other apps don't do still. And ChatGPT is kind of the only one that's been either at or very near the top from day one. So that's why I just send everybody to Chat GPT. So removing the AI boogeyman spookiness from it, which is to a degree, unfortunately, it's kind of just going to take time. It's kind of like cloud, where there was a lot of bad stuff with cloud and there were not secure ways to do cloud. And so people are like, well, cloud's not secure. And there was a little bit of just having to come around and get comfortable with it before you sort of just give in. AI's got some of that with it. But I would encourage you to be super first principles about this. Why, why are we concerned about it? What makes it fundamentally insecure? It is very, very important to have good like IT experts that are on your team that work with your firm. Don't take, for goodness sakes, don't take security advice from a YouTuber. Okay, let me, let me give you that first. But second, let me tell you, there's a whole lot of developers and even IT professionals who are so behind the curve on all of this stuff or unfortunately trying to shill you their own wares, like preying on the fear of all of this stuff. The reality is like AI models are like, how they use your Data is completely 100% well understood. But when it comes to any of the tech in your firm, you want to have a really clear process for how you vet it. And like, don't pretend you're the expert on that stuff. If it's me running my firm, you better believe I've got like my IT partner there and we've got a checklist. Like every single app goes through this process. At the end of the day they're signing off. It doesn't mean that there won't be some IT partners, unfortunately, that are like stuck in the past and also not making decisions and giving advice on like a very sort of first principles approach. But having trusted help like that has to be part of running your accounting firm. Now ChatGPT team, if we look at it and kind of look past like the, the AI boogeyman as we've talked about, unless your prompts are being trained into the model, then it's no different than any other cloud app. Stuff I like about it, you've got two factor authentication with chat GPT team, you've got SOC2 type 2 certification. There's a Number of leading practice management systems in our tool in our ecosystem that don't even offer that. But we happily put tax documents and all sorts of sensitive stuff in there. What could be better? And ChatGPT team for one, they don't yet give you a way to mandate two factor authentication across my team. So what I want to do is actually force every user to have to have it turned on. Still don't have a way to do that unfortunately but it can be solved with a single sign on platform like a Practice Protect. We've talked about this before Practice Protect which was actually acquired by RightWorks last year and they're now integrating it into their sort of IT platform. But you can still go out and buy Practice Protect. It is vastly better than it just a password manager. Basically it's a layer where your team signs in via that layer and then that thing logs into the app for you. And so at the Practice Protect layer I can enforce two factor authentication bit of a band aid solution. So I would love to see ChatGPT let me enforce to FA across all my users. Another thing that would be nice is like rolling deletion of data. So anything older than say 30 days you can set it to auto delete. They do that on their enterprise plan. But the ChatGPT Enterprise Plan is only available like the you can only start at teams of 150 people and up. So definitely room for improvement. But where it stands right now, if you look at like the chicken Parmesan like sort of standard apps inside of the accounting ecosystem, ChatGPT is probably above average with its security right now, specifically the team plan than the majority of apps in our space right now. Doesn't mean it can't be better, but it is not inherently insecure. Only other thing I think I wanted to cover we touched on it a bit. AI usage policies. These came around in a time when there wasn't a way to use ChatGPT that didn't train the model. And so there we would say like okay, non sensitive use case is good, sensitive use cases bad. And that's where the conversations around redaction and all that came up. And you still hear this a lot from people. And it's because there was a time when that was relevant, but it's no longer so you will still hear from people always redact, redact, redact, redact. And these days I don't think it's necessary for the same reason that I don't redact all my files when I put it into my practice management system. That was A talk track and a conversation that happened before we had ways to use AI models where all of your prompts wouldn't be trained into the model. These days, it is a security risk from the standpoint that it is one more app that has data in it. Right. Anytime you spin up a new app and you put client information into it, there's added risk there because it's another thing that somebody could get into that has stuff in it. I make pretty liberal use of temporary messages in ChatGPT. Is that what they call them? Yeah, temporary chats. So you can. You can enable this by default, and it makes every conversation ephemeral, so it immediately is deleted as soon as you're done having that conversation. I just use that by default unless I want that conversation to stick around. That was a lot of ground. Did we hit on everything? I think we did. I recognize the irony of telling you not to take security advice from a YouTuber. What I'm trying to override is all the security advice you're getting from, like, accountants and even people who call themselves AI influencers that just, like, lack a fundamental understanding of how it works and in the worst cases are, like, preying on that fear to, like, sell you something. In fact, there's a lot of people right now who are like, is it secure to use or not? Find out more at 11 and buy my course and then you'll find out. Maybe I'm a bad think boy for not putting all this information behind a paywall, but, man, like I said in the beginning, there's so much opportunity right now to do really cool stuff in your accounting firm from getting way faster answers for tax research. Like, every day I hear cool new use cases from accounting firms who have rolled this stuff out to their teams and are taking it seriously. Like, the best advice that I can give you, every member on your team needs to have the Chat GPT team plan. I can, I can go give a keynote and get, like, everybody frothing at the mouth about how this is a really big deal for civilization and things are going to change. They're like, yeah, no, we totally get it. And then I'm like, it's 30 bucks a month. And they're like, well, could you knock that down a bit? I'm like, gang, 30 bucks a month for the most powerful, like, software tool you've ever had. Access to my blueprint for modernizing your firm, getting your firm up to speed on this. Get Everybody on the ChatGPT team plan, but then have a weekly meeting where you get everyone Together you share use cases, what worked, what didn't work. The more contact points you can create where folks see what other people are doing, the faster they'll learn. This isn't a top down technology to learn, it is a bottom up technology to learn. And that starts by putting it in everyone's hands. Not everybody's going to take to it the same way. Hopefully you've got a couple energetic people on your team where they can kind of drag everyone else along and show the cool things they're doing to give other people ideas. But it really is as simple as that. You don't need one more webinar, one more podcast, one more training, one more course. You just need to start. That's the best way that you can learn how to use this stuff. Because AI assistants are like the most personal to you software we've ever had. Every other type of software, it's adopted top down. The big bosses tell you what to do. Every button is like you're trained on what every button does and all that. This is the opposite. It can do like anything you want to throw at it. Not always successfully, but the power is in the nuance here and the fact that it can solve very, very specific to you workflow problems. That's when AI assistants are at their best. Even if you're not bought in yet, I hope I at least resolved some some of these AI spookies for you because it is a really, really cool opportunity right now. If you're new here, welcome. We talk all things running accounting firms on this channel. Oftentimes that bumps up into AI and I'm out there in the wild hanging out with accountants like you. Upcoming events, I would love to see it. July 14th, I'm in Los Angeles. July 16th, Denver, Colorado. July 22nd, I'm in the big windy Chicago, Illinois. And July 25th, the Peach State, Atlanta, Georgia. Tickets for all those events@JasonOnFirms.com events would warm the cockles of my little soul if you made it out. Thanks for coming and hanging today. Questions, commentary, any of that stuff? Drop that stuff down in the comments and I'll see you in the next one.