428 Accounting Vendors are Using My Data with AI!

A whole bunch of accountants on LinkedIn recently got their ledgers in a bunch over an update to a terms of service that says basically we're allowed to use your data with whatever AI company we want to. We don't have to tell you who it is. We reserve the right to change that company anytime. People don't like it. They don't like the idea of their data being shared with third parties in a way that they don't have any control over. But at the same time, aren't we like all like poking our tech vendors with the stick, like telling them to like make an AI do a thing with AI, please make me more productive. And we're frustrated they aren't shipping stuff faster. How can those two things coexist? Because on the one hand we want to be responsible with our data. On the other hand, I don't want to get left in the dust if all my accounting tools aren't leaning into AI. Right? So today we're talking about this case specifically because it's worth digging into in a way that can just kind of generally educate and get a better understanding of, of how we should feel about this stuff going forward because it's becoming more and more of a thing. Let's do it. Foreign. So here's the update to Zero's terms of service that people don't like. By using this product. Here it is exactly in the words, a quotation. By using the product, you authorize Xero to process any data, including personal data relating to you or other individuals to the extent necessary to understand and manage your prompts and produce any outputs. You agree that Xero may share your data with third parties, including but not limited to large language models. This sharing is essential for the functionality of the product. We reserve the right to change third party providers at any time without prior notice to you, except where we have otherwise agreed. What's your gut, gut reaction to that? Probably a little bit uncomfortable. Like ah, really man. Like I don't really have any control over what they're doing with the data that I have in this software platform and the comments and there were a lot of them on this post as far as what I've seen are universally negative with the exception of Zero, who actually did hop in in the comments. But everybody's kind of mad. Everybody's throwing out governing body, this or that, the rule set, this or that. Surely this is in violation of it. So let's just, let's dig into all of the layers of this because there's a lot going on here. I mean, it's worth just calling out off the top that very much like accounting firms where we use a whole heap of different tools to make ourselves more productive. It's, I mean it's, it's naive to think the software companies are somehow running like an on premise server and like just using a hundred percent of their own tooling software companies the same way that accounting firms do. They use a whole collection of different tools and there's responsible ways to use those tools and irresponsible ways to use those tools. Now obviously around AI, this is a particularly hot button issue. And so much of this still comes back to early days AI where pretty much every AI assistant was training their models on your prompts. I mean, we had a good 12 months of spooky headlines after ChatGPT came out. From Samsung employees putting proprietary code into Chat GPT and being trained into the model, to people somehow retrieving like specific medical images from ChatGPT. All these really scary sounding things that made most people be like, I could never imagine using AI with client data. And I honestly, this is where 80 to 90% of accountants still are living is with, with the, the understanding that no, you could never put anything remotely sensitive into a large language model where. You know, most of the folks who tune into this podcast, you know, these days we have AI assistants that are every bit as secure, if not more secure than, than most of the generally accepted tools that we use in running small accounting firms. You know, the most notable one that most folks are leaning into is Chat GPT, specifically the team plan where no prompts are ever trained under the model. There's no way for anybody on your team to share conversations externally. And it comes with AICPA SoC2 Type 2 certification, which there's, I can count on like I think two fingers right now, the number of practice management systems in the space that have achieved that level of security. Yet we pump all sorts of data into our practice management systems, let alone other third party tools that are not as mature as the PMs. The short of it being many of the leading AI assistants, you know, from Google, from Microsoft, from OpenAI, they come with more security than most of the other tools that we use. Now this is also kind of falls into that Venn diagram of things you've heard around like when zoom zoom as in the video conferencing platform, they updated their terms of service maybe a year ago, basically saying we retain the right to train anything that comes up in your meetings into our AI models. And everybody's like, pump the brakes, bro. Like you, you're listening in on our meetings and training your models with them and they're like, oh, no, no, no, we're not, we're not really going to do that. That was, that was a typo. Our bad. We're gonna, we're gonna fix it to this other thing. But that is, that's a really important distinction actually, where oftentimes what makes people angry or the way that they, that actually this means that your data is being trained into an AI model. And that's actually a very, very different thing. What Zero is saying here is we will work with third party large language model providers and basically prompt those models and get outputs from those models. But it's not saying that it reserves the right to. At least not in this snippet that I've seen that they're able to train their own models on this data. Now there's a whole podcast episode that we could do that nobody would listen to. That is, that would go into whether that's actually a problematic thing or not. Because when you have large language models that are trained on billions of parameters, they generate a token at a time, which is like a word or a part of a word at a time. So with all the information that you have in a conversation, maybe you just wrote a paragraph and then you're asking it to respond, it uses that paragraph. And for all the, all of the data that it's been trained on, it takes all of the tokens in your paragraph, compares it to everything it's ever seen, and says probabilistically, what's the next word or half word that I should generate after what you gave it? And that's, that's like all these things do is generate a word or a half word at a time. It's pretty astonishing. This is such like a brute force approach to generating something that, that actually leads to pretty novel things that feel a little bit like intelligence. But for a specific thing to come up from a probability standpoint, it has to show up in the data set an astronomical number of. And maybe the best example of where this has actually played out is in the current New York Times lawsuit against OpenAI, where New York Times has said it's been able to prompt ChatGPT to output like, articles basically verbatim from the New York Times that would otherwise live behind a paywall. But if you dig into the nitty gritty of even that lawsuit that like, what they've presented are actually the free portions of articles that get reposted across thousands of sites the moment they go live on New York Times. So when the model is trained on all the data on the web and it sees this, you know, three paragraphs 10,000 times, there is a good likelihood that you can then like kind of sort of manually try to reproduce some of that stuff. But one thing going into there, like Steve's W2, a letter of financial statement, the notion that anything specific could ever be prompted out of that. When you think of the vastness of all of the data on the Internet that these models have been trained on, the notion that you could be like, hey Chatgpt, can you pull up Dougie Johnson's 2019 W2 dare? And then it generates one, one little token at a time. To somehow get to W2 like that is exceedingly unlikely, if not completely impossible. And that is, to be clear, the worst case scenario, something being trained into the model that shouldn't. I'm not saying that's okay. I'm just saying the whole notion that large language models are like a database you can retrieve things out of, not very realistic. But that's not even what zero is saying here, that they can use your data to train models. They're not saying that. They're saying we go out to third party large language model providers and you give us permission to do so. This episode is brought to you in part by the fine folks at Cloud Accountant Staffing who help you find talented accounts in the Philippines. One of the fastest growing trends in small firm running. Honestly. Ask around, ask in the comments, ask folks you meet on social media. A huge number of people have started doing this in the past year or two. Not just accounting firms, tax firms as well. Yes, I understand with US tax you got 7216 disclosures and that's not a reason to not offshore. It's a reason to disclose totally manageable ways to navigate that. Thousands of firms are doing so. World's getting smaller and honestly, it's never been harder to find folks that are excited about doing accountants onshore. Now does this change how you build your org chart out a little bit? It does, but I was, let's see, I was on a meeting with 22 other sub million dollar firms, tax firms specifically. And the prompt was, what's the best thing you've done this year? A couple of different people called out hiring in the Philippines for the first time and they were super happy with the people that they'd found Cloud Account staffing's background. They were an accounting firm, they started hiring themselves. Folks in the Philippines hired some recruiters over there to do that directly and now they help other firms do the same thing. Simple as that. That's something you want to explore? Have a conversation with me. Got nothing to lose. Learn more about cloud account staffing down in the show Notes. Oh, this episode is actually sponsored in part by, by Ramp, who I'm a Ramp user, I really am. My business actually runs on Ramp. It is a super, super solid spend management platform, which is a phrase that they made up because they can't call themselves a bank. So they're not a bank. But like spend hygiene and getting our clients onto better banking I think is, it's a really, really big opportunity right now. And it's one of those things where we don't necessarily like interject ourselves to say like, oh no, we should be opinionated with how you manage your money. But then you see like what a headache is to track down receipts across your team for just getting the month end close done. You see stuff like people passing around the office card where the numbers get stolen every nine months. Like gang, there is a, a solution for this, a way, way better way to do this stuff. We've done a couple of Demos on my YouTube channel now for Ramp. I would recommend you check those out because you may not even be thinking about like, I don't even know what's quote unquote, spend management or a good version of this. Looks like, give it a look, solves a whole bunch of headaches, is like the responsible thing to do. Like if you run a small business, I'm thinking like, put this information in front of your clients. If you do any type of bookkeeping, it can be a big old workflow hack to make your life easier too. Learn more about Ramp. Check out those videos on the YouTube channel or the link down in the show notes. Honestly, a really, really cool thing about the generative AI era is how it has actually democratized access to the world's most powerful AI models in a way that wasn't possible in past AI eras of past AI technology. In the past, you would hear companies like QuickBooks and Xero and all that say that we have this unbelievable advantage. Remember Big Data? Oh boy, big Data. They have this unbelievable advantage where they're going to invest X million dollars to train their own model on their proprietary data and they're going to do really, really cool things with it like automate classification and all that. Company said that for a decade and we kind of sort of got a version of that, but never a particularly impressive one. Right? But that was the age of like everyone having to build their own closed AI models. You think about IBM, Watson, stuff like that, that was people investing huge, huge amounts of money into training these models and their own infrastructure to do these things. And then GPT3, GPT 3.5 came out, the original model behind the scenes for chat GPT and overnight you had this public model and we'll define that. You had a public model that was light years better than anything anybody else had ever built in house. And so all of a sudden you had companies that invested millions of dollars in all these years into building their own models and they're like oh shoot, you know, for like a hundredth of a penny we can actually just go prompt that public model and get a way better response than we can get with our own models. And that was back in the GPT 3.5 days. So now, now you have companies investing tens to hundreds of millions of dollars to build massively powerful public models. And that's completely wiped out any notion of people training their own models. And so any company that's using the marketing vernacular, we've trained our own AI models to do xyz. They, they just haven't. The only element of that that they're kind of sort of is, is called fine tuning a model. And even that is not usually worthwhile because the investment to do quickly blown through in two months time when a new model comes out, the previous model's price gets cut by 90% and you can just straight up prompt the new model to do something that's even more powerful than what the old fine tuned model could do. So like people don't even bother with training their own models anymore. As evidenced by Xero using third party large language models. Intuit doing the exact same thing. They announced their partnership with OpenAI. And so really cool thing about this era of AI is behind the scenes these humongous companies with infinite resources, they're using the exact same models that you and I can use on ChatGPT. And I love that because the fear here is always that some giant megacorp is going to build some tech platform that you and I can never get access to and have this incredible advantage that we can't possibly compete with, right? And this age of generative AI has kind of has kind of like kicked that door down to say like no, we're going to have a few big players making amazing models that everybody, everybody has access to. And so it's evened the playing field now. I called those public models. Public models are those that like any of us can use. So you and I can go make an API call to OpenAI's servers and use their models the same way that Zero or QuickBooks would. When you're using ChatGPT, the same thing is happening behind the scenes. All Chat GPT is, is an app that on the back end makes API calls to OpenAI servers just like anybody else does. That is a public model that, that anybody can work with as opposed to say intuit developing their own model that just they use. That would be a private model. Now just because it's public doesn't mean that there's like a telephone wire somewhere with all of your prompts going by and, and everybody being able to see it go by and like see the stuff that you're prompting to. That's not what that means. Another source of confusion here is, and almost, I mean I see this almost every day on LinkedIn, people talking about OpenAI and saying like, I can't imagine ever putting client data into an open source AI model which OpenAI has not open sourced an AI model for a very, very long time. It is closed source, which basically just means you can't take the model and run it yourself. In no way does it speak to like the porousness of a model. Like it somehow absorbs the things that are bounced off of it. If it's open source like a, like a dish sponge that gets real stinky over time whether it's open source or closed source. It doesn't mean that. Take Deepseek R1, which was a super, super high profile open source model. Deepseek R1 super powerful model. I kind of turned the AI market on its head overnight. And you could get it in two flavors. You could download the model and the weights and all that and you could run it yourself on your own servers. It's not the sort of thing that's possible to run on your like, you know, actual computer. There are what are called distilled versions of the model, which is like, they make them way, way, way smaller so that they can fit on things as small as your phone, but they're vastly less powerful at that point. So to run the proper deep seq R1 model, it takes some pretty serious, like server architecture. And so the way that most of these companies make money is by hosting their model. So deep seek the company behind the R1 model. They also have a hosting service where you can make an API call to them and, and send your prompt to their model and you get a response back. Now that hosting service has been very problematic. There's a tremendous number of security issues with it. Prompts are trained into the model. All that data goes back to China, whether that matters to you or not. But that hosting service, it's problematic. The model itself though, is not the model itself any company can go out and run. If you've ever used Perplexity, it's like a really good AI search engine Perplexity AI, they use Deep SEQ running on their own servers. No security issues at all because they have total control over the model. They're running it themselves. Now if you compare that to what say OpenAI does with, you know, their GPT4O model, for example, that's a closed model. So we can't take that model and run it ourselves on our own server. So the only way that we can work with it is via OpenAI's hosting service. That is we send a prompt to their servers and we get the response back. And so if we want 0 or intuit or, or anybody else in the accounting vertical to build cool things with AI, they've got to do it the exact same way that anybody else does. Now that doesn't necessarily make it okay because then the devil is in the details. But when it comes to OpenAI specifically, and Zero didn't get into who they work with, when it comes to OpenAI specifically, API usage, like working with their API, their hosting service, to send prompts to the model and get responses back, that has like no prompts with their API have ever been trained into the model. In fact, the only situation where OpenAI ever trains prompts into the model has been in ChatGPT originally. Originally there was no like secure version and now with the personal non business plans, by default your prompts are trained into the model. That's why I always tell people to use the team plan. It's, it's like a pretty cost effective business plan. So we got to play by those rules. The only way we can work with OpenAI models is via their servers, the exact same way that Zero does. And for all the as. But I don't want to incite an angry horde of Australians, but I can tell you Australia, where Xero is the market leader by a mile. There's a whole lot of frustration with Xero. If you've seen the meme of the person holding the stick poking that, what are they poking is like a pile of dirt or something like that. I can't remember. They're poking it like do something. There are so many people that annoy that Zero hasn't moved faster on their product over the last five years and shipped really cool, really meaningful stuff or given us, you know, really interesting AI and advancements. So on the one hand the whole accounting industry is like poking them, being like, do something, do something, do something. And they're like, okay, like we're working with large language models and we now have like JAX and some other AI functionality that we're building into the product and then they just get, get like absolutely torched on social media for having to update the terms of service to say we have to work with third party large language models to be able to incorporate this stuff into our product. And I'm not saying like you should be excited about it by any means or that it's right. All I'll say is you can't have one, I guess, without the other. Now what alternatives does Zero have? Well, they could take an open source model and they could host it internally on their own service, own servers. Unfortunately this is usually more expensive actually than using some other hosting service, which, which is kind of surprising. You would think it'd be cheaper to do it yourself. But these other hosted models, because they have access to so much more server horsepower than a company like Zero does, and because they're actually kind of subsidizing it with investor money, that hosting, it's really, really cheap compared to, to hosting your own open source models. Open source is also usually just kind of perpetually behind the leading edge of what's possible. So open sour generally not going to be quite as powerful. Deep Seek R1 was kind of an exception to the rule where for a bit of time there it was probably as powerful as anything else, but already we have stuff that is more powerful than Deep SE Car one. And so even if you did say like, okay, we're going to, we're going to host all of our own models in house, we're going to use this. Like 45 days have passed and you already have models that you would be like licking your chops, looking at because they could do cool stuff that R1 cannot do. Now it is, I mean it sucks that like you don't get to know who they're working with behind the scenes. Part of the reality of this is who they work with is big business. Like when they announce a big partnership with a company. In fact, I think Zero even announced a partnership with. Who was it? I think it was AWS a couple of years ago. I think it may have been Google at some point. Intuit announced a partnership with OpenAI like 12 or 18 months ago, which basically just means they're using OpenAI's models like everybody else is. But behind the scenes, this is actually like big business, who you communicate that you partner with. Oftentimes companies will announce this partnership while using a whole bunch of other models from an entirely different vendor. And honestly, that's because different models are good at different things. So like there is no right now when it comes to AI, there really is no one stop shop that's the best at everything. In fact, now more than ever before, Google has some really, really good models. Xai's Grok 3 is really strong. There's still some argument for running your own open source models in house, but then OpenAI is still shipping really powerful stuff as well, as well as anthropic. If you've used Claude, they've got some super powerful models. So it's left to the software companies to then be like, okay, here are the different things that these models are capable of today. And some of them may work for this specific application that we're considering and some of them may not. And the hard thing is like, well, some of them may work 99% of the time, but then what are you doing the 1% of the time where they don't work? And it's put software companies in a super unenviable position to like have this thankless, loveless job where everybody's piling pressure on them to do more and ship faster and do some cool AI stuff and save me a bunch of time. But everyone's still scared of the AI boogeyman. And it's not totally unjustified because there are wrong ways to use AI. But too often the fact that there are wrong ways to use AI gets conflated into well, there's no good or no secure ways to AI when there absolutely are very secure ways to use AI. This episode is sponsored in part by Makers Hub. Hey, quick, quick win opportunity for your clients that do any sort of accounts payable. Even if you don't do bill pay, send them to this podcast sponsor, Makers Hub. You've heard me talk about Makers Hub before. We've run several Demos on the YouTube channel. Oh, in fact, bonus points if you send your client one of my YouTube videos. Okay, this isn't your average Jason Maker sub. They are leaning into AI vision models and large language models to enable a entirely different level of accounts payable automation. I've talked about this before. The, the, the single most disrupted part of the accounting profession right now, courtesy of AI is accounts payable and Makers Hub is a great snapshot of that done really well, really smartly. It's not one more platform to manage. It pushes all the data back into the QuickBooks file. So you still got a single source of truth. I have clients who this would have saved an unbelievable amount of time for. And this is when we were running Best of Breed Tech three or four years ago. For them. This would have taken that automation so much further. If you do bill pay as a service, you need to be looking at this because this totally changes your business model. But even, even if you don't, if you have clients that struggle with this and a lot of them do haven't checked out, Mega Sub super smart folks behind that, you're going to see him out at a lot of accounting conferences this year. Learn more Check out the link down in the Show Notes. Okay gang, this podcast is sponsored in part by Team up, the folks that help you find super talented accountants in the Philippines and hire them directly. Team up just launched a. What do they call it? A salary guide. A salary guide for hiring accountants in the Philippines. But it's so much more than a salary guy. You know all the embarrassing, ignorant questions we're afraid to ask about hiring people in the Philippines? How much does it cost to get this type of person? Can I even get, I don't know, a tax pro, an accounting manager, an executive assistant? Is that a thing? This guide, and it's going to be linked down in the Show Notes. It'll show you not just how to like get each of those different types of people, but what they'll cost, how to build out a job description, all of the questions that you have been too embarrassed to ask to find folks in Teamup, they put them into a guide that you can swipe. Get that one down at the link in the Show Notes. I don't like. I don't want to totally 100% come to the defense of software companies here because like, yeah, it absolutely does suck to have to sign off on a use policy that basically says like, yep, we're free to do pretty much anything we want with your data and share it with any third party that we want to. Because it's essential. It's essential to the product. That's just kind of a sucky spot to find yourself in. But if the shoe was on the other foot and a client says, hey, I was reading through your engagement letter. Why do you use Soraban? Or why do you use Carbon? I don't like that. I was doing some digging about that company, Carbon. Do you know that that company is based in Australia. I don't want Australians having access to my data. If a client could go through your software inventory and see all the things that were on there and they started giving you a hard time about it, what would your reaction be? I think my reaction would probably be, honestly, I can totally understand that. Like, I get it's a totally justified thing to be uncomfortable with and a little bit bent out of shape about, especially if land of lakes, if you look at the number of tools that we have to use. But at the end of the day, what am I going to tell my clients? I'm going to tell them, man, we've got like, we've got a process for how we vet these things. Like, we do have a due diligence process in the firm to ensure that these tools are managing that data responsibly. It's a very big factor in how we choose our tech. And you know, we, we work with a lot of other accountants too and IT groups and all these things to try to make sure that we're doing this in the most responsible way possible. Because data breaches do happen. Like we've had some pretty high profile data breaches already in 2025 and very common tools in the accounting software vertical. And so it's an unfortunate reality that that stuff does happen. But like, we are doing our very, very best to manage this as responsibly as possible. And if I'm talking to my client, I honestly don't know how to give them an answer any better than that. Like, I think that's the best that I can do. Corporations, big old companies, especially ones that you're mad about in regards to other things, it's hard to extend that same grace to them. Right? But one way or another, like this is increasingly becoming a thing because AI is becoming like generative. AI is just becoming a much more essential aspect of, of all of the tools that we use and will increasingly become so. And all of these tools, no matter what they tell you in the marketing behind the scenes, they're all using the same models. In fact, how this stuff works, it's, it's basically prompt engineering. So they're wrapping a prompt in like data specific to the application or data specific to how you're using. So let's say, let's, let's do a concrete example. A reporting app, let's say a reporting app has shipped a feature that will write an executive summary for you. Well, on the back end, what they're doing is they're taking some guidance from you or Some prompt that they've developed and they're injecting it with the real data from that actual company to say sales was this, this month, sales with this last month. They have kind of a prompt template and that template is injected with all sorts of actual data that will impact the response. So that when the model gets that prompt, the response is driven by the real data of that like very contextual client scenario. Right? And it is literally no different than you going out to Chat GPT if you had that full prompt and putting it into your Chat GPT. It is the exact identical thing. The only secret sauce here is all the things that you wrapped that prompt in. It's the data that is injected into it. And sometimes there's other like fancy handling around that. Like some apps that have like an open ended prompting box where you could type anything that you want in there. That's a really hard thing to control because they can kind of do anything and everything. And sometimes they'll do something dumb and then blame you. And so sometimes you will actually have like a secondary prompt that conditions the prompt that the user gave it or you will have this kind of other scaffolding that analyzes the prompt that you got in response to make sure that it's along the lines of what you want. And if it isn't, it'll go kind of through a review process and get like reprompted a different way to make it adhere to maybe the format that you want. So behind the scenes we're doing all these like weird gymnastics with large language models to get a very consistent response. But it's fundamentally the same thing that we are doing in Chat GPT as you're hustling that, trying to get something useful out of it. Now the only other way that you can kind of sort of keep identifying information out of a prompt to a large language model, even though if these prompts aren't trained into the model, I don't know the super necessary. But the other workaround that folks can try to do here is called substitution. So basically if you think about a prompt, let's go back to that executive summary prompt. There's a whole bunch of information there where, where if that prompt isn't tied back to anyone identifying it kind of doesn't matter. Like your sales figures, your financial statements, all that stuff. Like if it goes to, if it goes to somebody's mailbox, you send that out. But it has no context about who it is, not that big of a deal. There's a risk of it becoming a Bigger deal if you know there's something associated with that document that then can be tied to ABC Corp. And then all the, all the connected documents, you now realize all those things are ABC Corp. But a redacted version of information, it's better than the unredacted version of information. But a step further than that is called substitution. And that's where you take anything identifying oftentimes numbers, that sort of thing. As much as possible you substitute that little bit of text inside of a prompt. So if you know John's name was in the prompt, we could substitute John's name with Craig. And we send that thing off and we have that like key value pair on our own server before it gets sent off to know that when it says Craig back, what it really means is John. So then it gives you the response to that prompt. You change all the Craigs back to John. And now the model didn't know that, that Craig was really John. And there's people that have like tried to do versions of this and where it kind of runs out of Steam is that like no version of this is 100% perfect at dynamically picking up what those identifiers are. The other thing is sometimes those details are actually very important to the model doing something useful with it. So if it is say the sales financial results one month versus another month, the model needs to know what they are to be able to give you something helpful back. A few companies have said like, oh, we devised our own way to ensure that no secure information is ever sent to the large language model. And honestly it is a marketing thing to attract like the folks that are super spooked by this. It is some version of subtraction. But your usage of the language model is fundamentally limited when all of those details have to be stripped from the prompt itself. Which is actually why Xero says in their terms of service that you're authorizing them to process any data to the extent necessary to understand and manage your prompts and produce outputs. So sometimes those identifying details are necessary. It is a hard problem. I don't want to make it sound like it isn't. Um, it's. I mean there's so many crossovers to like the cloud here and discomforts with the cloud and, and similar to AI, there were all sorts of bad ways to do the cloud and bad things that came from the cloud and bad actors and people who got it wrong who are well meaning. But it, it is like it, it's also the future. So like we have to be smart about what are the what are the intelligent ways to do this? If I've hitched my wagon to zero for all things, for processing payroll, for all these different things and secure information that we're putting into that platform, I don't personally myself have concerns if they're like, okay, we're gonna do the AI thing now, I'm actually pretty excited about that because similar to a client coming to me who has no idea about the tools that I use, I'm like, relying on them, having a degree of faith in me to do all that, navigate all that responsibly. And there will be people who don't do it responsibly and companies who do it abusively, unfortunately, just like there are bad actor firms that will do some really awful stuff. But I don't myself like my firm, I don't want to take the heat for what other bad actors may do with that. And the worst timeline here, the worst timeline here is if we have an accounting tech vertical that's so afraid to lean into AI that everything else passes us by, all the general tools then become more powerful than our industry tools. That is the worst case scenario. And every time we drag software companies through the mud, all the other companies see it. Like, there's a tremendous amount of fear and trepidation in a more mature company, like a practice management system shipping AI features because you have existing users that will freak out, that will leave the platform. You will have stuff that sometimes goes sideways, like Zero's Jax AI Assistant. It's had some issues where it's returned some kind of weird stuff, and people will post that stuff on social media and absolutely just drag them through the mud. But it's like, dang, like, what's the alternative? Like, yes, there is a spectrum. And like, did we just ship something that sucks or did we just ship something responsibly that's actually good? And it's not to say that just shipping is the answer because you want to actually do it right. But people are scared. And, you know, we've been hammering on this podcast for a long time, really talking about today ready AI applications that we want more of our vendors to use. And I know a lot of you go and run to your vendors and you'll even link them the podcast and be like, man, this would be so good if we could. If we could get this and we could get that. And Jason says, it's super easy to do. Why haven't you built it yet? And oftentimes the reality is it's complex. It's like the The PR of it all and the fears that folks have when you outwardly brand something as being like an AI feature, like that's still a really messy thing that that makes some folks uncomfortable, unfortunately. And as much as you know, I am a believer that AI is just kind of this inevitability of where things are headed. I still don't want to be a tech apologist and I don't want to discount just how uncomfortable this is and how it's like I may not super love the idea that I'm like handing all this stuff over and just kind of giving them full permission to go and use my stuff with different AI models. So I can also totally understand that end of the spectrum. Just kind of a hard spot that we're having to live in right now. As with all things AI, the solution is to get comfortable with it. There's a man, we may do an episode on this. There's something like 25% of job applications now in the US want AI proficiency. I'm not seeing this very much yet in the accounting space, but AI proficiency just becoming like a base level skill to be a productive knowledge worker. Feel like in 2025 we kind of hit a tipping point where like I'm hearing from so many firms are using ChatGPT team for the first time and having really good luck with it. Deep Research is doing some just amazing things that we haven't really gotten to dig into much yet. But the path to like getting comfortable with AI is just starting and a very comfortable, very safe place to start. ChatGPT the team plan. I wish I had some sort of affiliate on that thing. Man, the number of firms I have never shilled so hard for a product in all my years. But it's very, very good. I use ChatGPT every single day because accountants love their monitors and I don't want you leaving your email up on a monitor. My tippy, top tip, leave ChatGPT up on a monitor. Cover up that email. Don't let it distract you. Use Chappy ChatGPT for more day to day stuff. Like I find ways every single day for it to save me time and I don't know a way to get comfortable with it without just using it. And the reality is, behind the scenes, when you're using ChatGPT, it's working with those OpenAI servers the exact same way your accounting software is. If there's anything I missed here or like even technical details, if you're, if you're a mega geek, drop that thing in the comments. If you have, like an embarrassingly simple question, you think that something you've always wondered about AI, that's really valuable, too, if you have that question. A hundred other people have that question. Drop that down in the comments and I'll hop in and answer it if I can. It's important conversation to have, like to try to just move this whole conversation forward, help educate folks. Educating them does not necessarily mean that they're going to be happy about it, and that's okay, too. But it's a conversation that we we just got to keep having. Gang, that's it for today. Thanks for coming and hanging. I'll see you in the next one. It.