359 A Peek Into The Future of AI Accounting Tech

Man, accountants are not excited enough for AI. And I completely get it because we, like, look around at the products we use. What's the last cool AI feature where you're like, wow, that really saved my bacon. Thanks. I. I don't know that I could put my finger on it. We're getting like, incremental stuff. Some stuff around tax intake, some stuff around bill pay that's pretty novel. But for most of us, we're still like, my body is ready, please. Like, just the sooner it can come, the better. And I've had various diatribes about this recently on the podcast and today I just want to go full nerd on a whole bunch of things that we're ready for that are possible with today's tech. We're going to, we're going to force this conversation on our vendors to say, I want this and I want it now. Because there's so many super cool things around the corner. In fact, I shared a bunch of these on social media and so many people came through and said, we're about to ship this. We're working on this. Like, go down the list. Carbon can. Like, there were a whole bunch of people that said, like, yep, bunch of these we'll have soon. So today we're going to dig in and just, let's, we're just going to get excited about what's to come because thanks to AI it, it'll go better next year. Okay. So I have a, I actually have a software company facing newsletter. I talk with software companies a ton. This phone, this, you bet. This phone's got the, the, the digits of the majority of the CEOs, of the biggest tools in our space. And so talk to software people all the time. And I know what folks have in the works. I've also done a lot of like, very technical AI advising with software companies to understand like the gummings of, of how you build these things and what actually works in production and what doesn't. So, so this is real legit stuff. So I'm going to go through. I don't know that I'm going to go through every single one. I'm going to go through some of my favorite things that are in the pipeline. First up, I love this. I love this one. So much. Scope reviews. Okay. Does a platform have access to the service terms of my client engagement? So think about where that lives right now. If you do engagement letters in your practice management system, that would have it. If you use a standalone proposal platform like Ignition or Anchor, that would have the service terms but think about how we review, say how a project went. Maybe that's a realization report, maybe that's, you know, did we hit those target deadlines, number of hours, whatever that report looks like for you, you know, what a really valuable addition to that report would be? Does it look like we did anything that was out of scope? And I'll tell you right now, in fact, if you have ChatGPT, you could go test this today if you wanted to. The O1 preview model, which is currently the best model that we have in chat GPT. If I put the email interactions I had with that client, the files we created for that client this last month, if I put that and I put our Service terms into 01 Preview, honestly, you could even do this with inferior models and I think it would be pretty good right now. It would actually give you an intelligent, informed response to say, here are a few things that might have been out of scope this month. And boy, especially if you run a team, there's nothing more frustrating than getting six months down the road and realizing we've been doing what for the client. And we didn't have any conversation about scope. This is another argument for like production based compensation for your, your team members to be actually incentivized to have scope conversations with clients and renegotiate fees and all that stuff. A firm owner told me that they were working with a startup and the startup raised like $40 million or something and got a wire that was a 30, $40 million deposit into a bank account. The bookkeeper booked it and just kept on chugging along through the month end close. And six months later they find out this company raised a Series A or something like that. Is that ultimately entirely the bookkeeper's fault? You could fight that one a bunch of different ways, but this is 100% a situation where all artificial intelligence could lend a second set of eyes to help us out. Next feature. How about file naming conventions? Do you have a system for this in your firm? If you haven't like seen a good version of this before? What we had on my firm was basically like classes of documents. And so you're going to have an explicit list of. For these 20 types of documents, anytime one of them is in our file system, it needs to be saved with this like naming convention. So financial statements had a naming convention. Tax returns did certain types of work papers had naming conventions. If there were bank statements or loan statements or whatever they were, a general ledger detail, they all had to be named specific ways. And getting a bunch of humans to Follow the same set of rules for this is admittedly challenging. But you know what would do this really well? AI in fact, like, in a perfect world, I would love it if the platform would give me a way to dictate what the naming convention was for a certain set of, like, types of docs. But even if the platform said, we're just going to do it this way, like, we have our own system and for these 200 types of docs, we're going to name it in a standardized way, honestly, I would jump out of the way. I would click my little heels with glee and I would say, great. Now I have to look never look at a, a poorly named like file again and wonder what it actually was. Particularly if you think about contexts where like, a client uploads stuff from the smartphone and it comes through the file name, that's just completely meaningless. Now you have to open it and I have to rename it to something that's helpful. There's a bunch of different versions of this. There's also a version of this where labeling may actually be just as helpful, if not more helpful. If you have like an explicit set of labels, like the tax year, the type of document it is, whatever that is, AI can, can label that stuff very accurately on its own. But today our file management systems, they are completely human led, and in most cases, just like a folder structure is absolutely better ways to do that right now. Next up, if you're going to write something for me, AI, how about try matching my tone? So increasingly, practice management systems are including like email clients in the platform because it's such a big part of work. And if it can help me by surfacing relevant context when a client asks for last year's tax return, why did I have this or that in my financial statements? Like, it totally makes sense that practice management systems are getting more and more into email because they can help us answer those questions, right? A few platforms now, you can have it draft an email based on like a short prompt. So tell them, yes, I'm happy to meet this Wednesday and thanks for sending the cookies. And a draft's a little thing. And honestly right now I don't know that anybody actually uses that because the output feels so stale, like something that came just out of ChatGPT. Because it did, they're calling the same GPT models that everybody else is. But a version of that that's way, way better is one that has a prompt that will better target it to my tone of voice. So take a practice management system, for example. They Have a kajillion examples of the ways that I email my clients and the tone of voice that I use in those emails. From those emails, develop a three sentence prompt that describes my tone of voice and anytime I ask it to generate something prompt, pull that prompt in. So on its own, the platform is going to, going to have AI analyze a bunch of my writing and say write a three sentence explanation of this person's tone. That's then like a behind the scenes user setting. Or maybe it's something it'll actually let me tinker with. Like maybe it auto generates a starting point, but then if I want, I can change it because maybe I actually, maybe the tone I want it to generate for me is a little more aspirational, like maybe a different tone than I use naturally, maybe more professional, maybe less professional. But ultimately the useful version of the tool that generates writing for me is not the one generating the best writing. It's the one that generates the writing that feels most true to me. Correct in this case is it's like me. And this is not a hard thing to build into a platform for it to always wrap that request for write an email based on this prompt to wrap that in a prompt that describes the tone of voice to use for you. Now you could go actually wild with this. Like, you could make these tone prompts even specific to the people that you talked to. Because when I talk to Jeff, like I'm using curse words, I'm. I talk to him very differently than when I talk with Belinda. And so if you want this writing feature to feel magic, make it contextual, like that's something that is one more API call to OpenAI to write a little prompt that's specific to this client. Not that hard to do, but greatly increases like the kind of magic that feels like when I'm responding to Jeff and what AI generates is, hey, Face, are you going to send that document to me or not? Okay. Up next, transcript support. Accountants are amassing mountains of meeting transcripts right now, as you should. AI meeting transcripts need to be, need to be creepily lurking in every single meeting that you have right now because you will never regret having those transcripts. Imagine if you had transcripts for the last 10 years of meetings and they were searchable and you could be like, did he tell me he was gonna claim his daughter last year? Was that the year before? Did we talk about the rental? My colleague just met with that client yesterday. What did they say? Did they talk about this thing in the meeting? Or do I need to Ask them Transcripts, so valuable you will never regret having a meeting transcript. And honestly, the only pushback people have is like, I don't like the notion of this thing recording what I'm saying. And buddy, if you got a problem with that, then it's probably not a conversation I want to be involved in to begin with. Like, I'm not only ever going to call you on my burner phone, that's just not how this is going to work. So a growing number of accountants are using meeting transcripts and we're piling them up and we've got them and it's better than nothing. And you can go dig through them if you need them, but give me somewhere to put them. Like every tool in the accounting ecosystem is just pretending they don't exist right now. Like, I could put the file into my file system, but these file systems don't do like deep search into the contents of the files. Like it's doing a very surface, surface level search, which actually gets me to the next feature that I have is vastly improved, like deeper search. So AI has made possible searching vast amounts of information via. It's called Retrieval Augmented Generation, or rag. Basically you can have just about an infinite amount of information. And what it does is it will take a query and it will pull the chunks that are most relevant to the query. So you could have a million pages of documents. But I'm asking, what was Thomas Jefferson's favorite flavor of ice cream? Now within those documents, maybe there's 15 pages that are about Thomas Jefferson or about ice cream or somehow tangentially, 10 tangential, you know what I mean? Kind of sort of related to the question what Retrieval Augmented Generation does is it can look through those million docs insanely quickly, pull just the relevant chunks, and then you throw all those chunks to a language model. And the way it actually works is it says using all of this authoritative, like context, the, you know, the 10 pages that it pulled that related to Thomas Jefferson and ice cream. And with this context, answer this question. That's Retrieval Augmented Generation. And if you think about our practice management systems, what are the sources of information that they have? All the files, all the contents of the files, all of my organizational email history, my client's accounting ledger, a growing number of PMs are integrating into the, into the ledger. And they should. Meeting transcripts, calendar events, projects, tasks. There are so many things these systems have access to. Wouldn't it be nice if search was actually intelligent, if you could search across different types of information from meeting transcripts to tasks like this one, search to rule them all, where I could literally type in the name of that client's rental property. And every single workpaper, every single tax return, like every meeting transcript where it was discussed, every email, it all actually shows up so that I don't have to go looking and emptying different places to try to find it. It's a thing, it's something we totally have today. There are all sorts of examples of this in other platforms. Like ClickUp's doing a good version of this now. I think it's actually something that we're gonna see be vastly improved by platforms very soon. Today's episode is sponsored in part by liveflow. They just shipped liveflow dashboards. And if you're looking for a little way to gussy up the monthly reporting that you do for your clients, this is pretty nice. They like modular. You put your own commentary in, not a, like giant reporting suite, not something the client has to go log into. In fact, if you're looking for like how to wade into more viable advisory conversations with clients, oftentimes you can only get there if you're giving them information that they can understand and sending them a couple of PDFs of financial statements that is not likely to lead to advisory conversations because they don't know how to make that stuff actionable. The more that they can understand why what you're sending them, the more likely it will lead to valuable conversations. LifeLow's got a bunch of templates you can swipe to get started. Like even a P and L dashboard, that's pretty cool. A daily cash movement report. A lot of stuff that if your client saw it, they'd be like, oh, I would like to have that. And from there you then go from just churning out financials to being irreplaceable their buddy. And these days we all need another friend. We all need more buddies, right? I think so. Everybody else, Everyone in the room's nodding their head in agreement. Learn more about LifeLow. Check out the link down in the show notes. This episode is sponsored in part by Client Hub, who I recently named one of my top practice management system recommendations heading into 2025. But according to Client Hub, when it comes to picking a pm, it comes down to the four C's capabilities, client experience, clarity and community. Today we're talking client experience. The biggest blocker to like getting stuff out the door with clients is waiting on them. Them not following through on what I need, which is what makes client have a really unique client client experience. It's basically Like a two sided task management system. You're doing your work over here, client's doing their work over there. It actually looks largely the same the two sides of that system. This isn't like a portal that they bolted onto a task management system after the fact, which is what most PMs are. Client up was built for the client experience from the ground up. Since that's what's going to make you more productive. A better client experience. If you've been unhappy with that side of the system systems that you use today, Client app might be a great fit for you. Learn more about that one. Check out the link down in the show notes. Okay. Generative Draft Email replies We talked a bit about email. If somebody can nail Generative email, like that is going to change so much and it's, it's inevitable that we get there. Like it's genuinely inevitable that we get to a point where we don't have to write most of our emails where it will draft something and you skim it and you're like, yep, that sounds like me. That's what I would have said, come to go. And then there's a point beyond that where there's a confidence level where it's like, I've gotten this exact Type of email 10 times in the past and it's always fine. So I'm just going to send it off. Like that is inevitably where we're going. The point where it will automatically send emails for you and you're like, well, what's that? Seems awful. What does that mean? Like my, my AI agent's going to be emailing somebody else, but then their AI agent responds to me automatically and they're just, it ends up being a whole bunch of AIs just talking to each other. And the reality is like, yeah, that is where we're going. You know how many machines are talking to each other right now? Every time you issue a bill out of bill.com and it pushes that data over to the QuickBooks ledger like it is just another interface. And if it saves me having to harass a client for a W2, if I can email that client and ask for that thing and their AI agent responds right away and gives that back to me. That sounds great to me. Now obviously there's a whole bunch of boundaries we have to consider there, right? Like, should they be emailing me W2? Probably not. Do I know who this person is? Like, do they have authorization to, to get this stuff from me automatically? So there's a whole bunch of stuff like that to figure out for sure. But eventually we're gonna get to the point where we get an email from a client and this is like, honestly I thought we would have a better version of this than we do right now, but it's coming soon. You get an email from a client, they're asking for a copy of the prior tax return and it drafts the response and attaches the prior tax return. Like that is not rocket surgery, but even more complex things like the banker was asking why net income was down last year. Think about kind of the multi steps of logic that go into that. Pulling financial statements, looking into like follow on detail. This is stuff that we're seeing from the O1 model now from OpenAI for example. It excels at sort of multi thought, like multi level reasoning to get to the bottom of a problem. So the future of email is changing. Turning up to the inbox with something that's pretty good already or at the very least all the context I need to build my own answer. Now the best version of this that I've seen in the wild thus far is by Superhuman. So Superhuman is an email client. It's what I use. I don't know that I recommend you use it because in a perfect world, your email is super integrated in your workflow. If you're using Outlook today, I would use Superhuman instead. If you're using your PM for email, I would stick with your pm. But what Superhuman does for AI replies is for every single email that you receive, it drafts three potential responses. And so when you get that email from a client, it may say, hey, are you free to meet with me next Monday? And there will be three buttons below that. One says yes, one says no, and then one says, you know something else, something in between, or something like that. But if I read that my answer is going to be yes, all I do is I click that yes button and it has already drafted a yes response. I click that button, it puts that response like in the body. And then all I have to do is proofread it and hit send and it is matching your tone of voice. Like we talked about earlier, it's learning from things like do you usually share a calendly link in this situation? And it's pretty wild if you think about this. It actually means I should never have to write an email from scratch again, right? Because one of those responses it's suggesting is probably gonna be close to what I'm gonna say in most situations. I click that and I'm probably 80% of the way there. And as it gets Better at mimicking my tone and understanding I other context. It's going to get 90% of the way there. That's a really cool today. Possible version of generative email replies that nobody shipped yet, but I hope we see soon. Requests. We make a request to a client, which we're increasingly using software platforms to do, right? Not just sending emails. We make a request to a client, we need this doc. But what if that doc's already in the file system? What if that doc was discussed in a meeting transcript when my client talked with a colleague of mine yesterday? Our request platforms don't do anything to keep us from making a dumb request. An embarrassing request where the client comes back and they say, you already have that. And sometimes we do, sometimes we don't. And even before we make the request, you probably just burned 20 minutes looking anywhere and everywhere to make sure that you didn't have the file. Because you're going to feel like a total idiot if you ask for it. And you did, right? So like the system double checking your request before it goes off and being like, hey, I saw this. Then the other thing, could these resolve the request actually before it goes out? That not only reduces the risk that I end up looking like a dummy, it can actually replace the time I would otherwise spend trying to look for the doc. If I just go straight to making the request and it says, review these three documents, because this might be what you're looking for. And if that saves me from having to go out and do that myself, boy, is that a big time saver, right? Similarly, how about resolving requests that are already outstanding? So if I have a meeting with a client and the transcript goes into that platform and it resolves three outstanding requests, maybe I'm the big boss and my team is blocked by those requests. As soon as I get out of that meeting, I want those requests to be resolved with the information from the transcript. Bang, bang, bang, that stuff goes off, my team gets notified, they can keep working. Similarly, if I have a recurring request that goes out on the first of the month, for example, say we have to request a statement from from the client on the 1st of every month and they get a request automatically. If the client emailed me that statement two days prior, that request is still going to go out and I'm going to look like a dummy unless I remember to turn it off, which I won't. So being able to resolve like upcoming automated requests, that would be really nice as well. Isn't this like, Ian excited yet? Like, this is all really good stuff. That saves us a load of time and makes us look smart, which I think that's ultimately all of our goals. Right. All right, this is biggie buckle up. AI driven checklists. I was real hot on this about six months ago when we got the last version of GPT4.0. And this just got way more real with the 01 preview model that we now have in ChatGPT and the full 01 model, which we should be getting any day, which is supposedly another big step up but beyond 01 preview. So if you recall previously on JSON on firms, we put this new model to the test reviewing a set of financials and it went through a balance sheet and pulled all the same potential issues that I did. Like, it's, it's pretty darn good and I would argue today a very useful tool for the preparer to do a self review before it goes to the big boss. Now, AI Driven Checklists. The way I see this panning out is we have, you know, there's a whole bunch of what we do that can't be completed via checklists, but can at least be improved. In fact, I'm right now I'm listening to the Checklist Manifesto. I've had it in my Audible library for years. I'm finally getting back around to it. If you've listened to it. What I wasn't ready for was the number of terrifying medical stories that are in there. But what I'm really enjoying about the Checklist Manifesto, like it's going through some examples of where checklist, just basic checklists and medical applications have vastly improved all sorts of outcomes and now it's looking at kind of other industries and we all have the same blockers. Like doctors are actually a great analogy for the biggest brain accountants out there, I think because they're almost offended by the notion of needing a checklist. And they're like, what we do is so infinitely complex that it could never be captured by a checklist. Like they're literally like fixing human bodies and all the different things that can go wrong and the different ways people can get stabbed and punched and fall off of things. And you're telling me a checklist, like that's going to be my solution to improve outcomes. And the answer is, is actually yes. Not that 100% of things can be checklist defied, but the fact that anything that can benefit from a checklist should have a checklist. And so what I'm most enjoying about this book is how it's kind of breaking down all those different barriers and also Interestingly, shining a light on how we all think that our job is the hardest job that there is. Like they're talking about ER docs, but they're also going to people who oversee like skyscraper projects where you'll have 400 people working on this building site a day and somebody is making sure that they're all doing the things that will in the end will result in a building that can stand up on its own. Unbelievably complex things, trying to reduce them to a spreadsheet wherever possible. This is absolutely possible within our profession to be done much, much more than we do now. But I would argue the reason that we don't is if we take this to the extreme, it's actually like almost inhumane to think of the volume of checklists that you would put in front of a human. But if an AI were capable of completing that checklist, oh buddy, what does that change? I love the idea of having a hundred item checklist to make sure that this medical deduction is deductible and we got everything that we ought to have for it. But would you ever in good faith send a human being through a hundred point checklist to verify all the stuff that you want to verify there? Probably not. Right. So the testing I did with the last model from OpenAI, this is before we had O1. I was testing things, very simple things like we, we put all of our financials out on cash basis. So it would check. Is the, is the report cash basis? Yes, it could do that. GPT4, even, even the previous model can do that very reliably. You give it the financial statement, it can be like you yes or no, is it a comparative financial statement? We always did current period to prior period. It could, it could double check that. Yes or no, do total assets equal total liabilities and total equity? It could check that. Shockingly, our ledger software does not always do that reliably, but the language model could confirm that. Yes. So those are like three very basic examples of checklist items where if you could snap your finger and an AI would take off and go check all those things for you, you would 100% do that. And the beauty of this is we have had all sorts of quality assurance tools out there where you plug in the ledger and it runs all these checks for you. But the reason why I think none of them have taken off in a very mainstream way is there's just no open ended way to do that. Where I wanted to specifically look into this and specifically look into that. And if you have your own way. And there's like 500 things that it checks on its own. I'm, I'm fine with that too, but I want to then add to that list and have it check X, Y and Z and language models. The way that you explain to a language model to check this thing for you is the exact same thing that you would tell to a human. So we are all capable of being quote unquote coders in this regard. And so if the previous 4O model could do this really well, the O1 model, O1 preview that we have right now, like, it can absolutely do this. The biggest blocker right now, honestly, is we just don't have a great user interface to do that. Like, if I want to run a hundred prompts through this one right now on O on preview, it might get expensive. I say expensive like this is in relative terms. But the cost of language models are going down exponentially every single year. They're getting way, way faster. But I'm talking about like for less than a dollar of LLM usage, you could have it run 300 checks in less than five minutes. But ultimately you have control over what those checks are and they become part of your quality assurance process and you have total control over them. Like, we're, like we're there. Like the models are good enough today. We just arguably don't have a well productized way of doing this. And so if I were to tell you right now, and you believed this, that 12 months from now, nobody would ever ask a human accountant to complete a checklist again. Nobody would ever ask you to go through these 30, 40, like checklist items. And we've got an AI for that. It'll do that fine. If you truly believed that, boy, like, what does that change with how we approach our work today? We probably start thinking a lot more about systems. Probably start thinking a lot more about what our quality assurance process ought to be and what those checklists, checklist items could be. Because I think we're there. Like the models are there. It just hasn't been productized in a compelling enough way. And the counter argument here is like, oh, they're just going to build it in the QuickBooks, are going to build it into this or that. The reality is they're going to build in something, a version of this. But we've always had quality assurance tools that will do generic versions of it. And the reason we haven't used it is they couldn't be tailored, they couldn't be customized to look into the things that we want them to look into this is the power of language models where we literally just tell it the same way that we tell a user with a paragraph describing, look into this and it does it and it produces a work paper to show its work. Now, I've actually been pushing a number of platforms to allow for semantic quality assurance for quite a while. Let me put in a whole list of prompts that are things that I want you to check on each, like every single month, end close and then just show me the result or create a work paper for it. We're talking about checklists here. But this approach to quality assurance, this is ultimately how everything gets reviewed long term. So if you think about a work paper management platform, if a project is otherwise done, how do you have it? Check a whole bunch of things, how do you have it? Make sure that the pages are annotated the way that they're supposed to be annotated for your firm. Ensure that this value matches that value. Like, this is it. It's having language models review A and B according to these instructions and say like, did it do it? And if you want to have a play with this, it's, it's fascinating. Try uploading a couple documents or copy pasting some text from two things into ChatGPT, giving it that instruction, like that, that one line, single checklist item and say, ensure that taxable Oregon state income only includes blah, blah, blah, that was actually deducted last year. Like whatever that super contextual version is for you. Give it a try and you'll be amazed at like how well the models can already do this. Super exciting times. Today's episode is sponsored in part by Fin Daily. Man, there are like so many powerful tools now that can automate some adjacent services to what we do. I'm a big advocate of not being in the business of just selling more bookkeeping or just selling more tax prep, because that's usually where we're most constrained. It's the work that's the hardest hire for. Rather find a specific type of client with adjacent needs where we can build adjacent services that don't have this like commoditized value that's easier to hire for, that's easier to automate. Fin Daily. Fantastic example of just that. It's actually an app that I founded Later Soldiers that was built just to automate the cash advisory services I was providing to my clients. We started doing bookkeeping, then we got into bill pay. Then we learned that if we do bill pay, we had a great understanding of their cash. And this cash reporting made all that other like work worth it. We can make more money on the cash reporting than all the other stuff combined. And it was way easier to automate. I did that with Bin Daily. It's a super cool tool if you haven't seen it before. Learn more about that one at end the the link in the show Notes this episode is sponsored in part by Mercury. Hey, you know where my money's sitting right now? It's sitting in a Mercury bank account. Look at me chilling the dog food that I eat. Mercury today is loved by over 200,000 customers. Does business banking that is transformative. Corporate debit cards, credit cards for expense management, free wire transfers, treasury yielding over 5% interest on idle funds. But Mercury for accountants, that's you lean in. It's their new accounting partner program which wants to get you your time back and drive client value. This is cool. They got this thing called Panorama. You can actually see a whole bunch of client accounts in a single view so you don't have to like drill into all these different things or log in and log out of umpteen different accounts. Panorama. That's right. That's. What a great name. Plus they get all the cool features that you do if you use Mercury banking Bill pay, custom approvals, invoicing, document upload, GL code mapping, reliable integrations with QBO 0 Netsuite, everything you need to make those accounting workflows sing. Learn more about Mercury. Head over to mercury.comaccountants to learn more about the Accountant channel. Or check out the link down in the video, the show notes, the words under the thing. Another pretty wild application. The notion that clients could be able to log into their client portal, like probably with your practice management system, and chat with an AI AI assistant that could see into that client's context. So their tax return filings, their accounting ledger, maybe past meeting transcripts where they've, they've talked with you. Maybe that AI assistant can see into all of that context. Kind of zooming out. If we think about the future of quote unquote advisors and the notion that AI advisors are inevitably going to get pretty darn compelling. Like they're going to be very good chatgpt today. Like if, if you showed this to somebody five or 10 years ago, they'd be like, well that's, that's wild. That's, that's unbelievable. I mean, 01 Preview is solving problems at beyond a PhD level. And if you had told somebody 36 months ago, we would have AI models where you could just chuck something into it with no other context and it would complete those problems at above a PhD level. You know what your reaction would be? Well, I'm not at a PhD level, so what does that mean for me? Like we would be beside ourselves if we saw like what we actually have today. And so in all likelihood, AI advisors are probably going to be pretty darn good, but they're only ever going to be as good as the context that they have. And today, if you think about what advantage accountants have is we are the only ones that are the keepers of all of the relevant context for the client. Tax returns, the ledger, past email, conversations, meeting transcripts, we have all that stuff. So who is in the position to stand up the very best AI advisor for our CL clients today? We are. Or more accurately, the platforms we use are. So I'm not in love with the idea of like a client sitting up in bed at night and, and developing a relationship with my client portal. Like I'm not more into that than you are. I might be a little more into that than you are, but that kind of just feels like the inevitable road that we are headed down as more and more people talk with AI advisors. Like we kind of have to have something to say. There's because our version that has all that context is going to be better than, than the TurboTax version that only sees their tax return. It's going to be better than the QuickBooks version that only sees their ledger data. Like we have way more context, so we ought to be able to stand up an assistant that is more helpful. Now what, what becomes hard then is what if it advises them in a way that I don't? And like to that end, I'll throw one more feature at you here that, that's related to this. Oh no, I got another one after that. Okay, two more. But on the subject of that client facing AI assistant, what we don't have still is advisor in the loop assistance. Because what I want is the AI assistant that can serve as the first line of defense for common questions for me because 80% of them are low hanging fruit, not hard to answer. I want my AI assistant to get those answers to clients immediately so they don't have to wait for me and so that I can save my big brain for the other 20%, the stuff that actually requires me. So give me the AI assistant that at a certain like confidence threshold that I can set, will hand off to me. Now this is already a thing in other tools. In fact, let's link this in the show notes. Our JSON bot is built on Delphi D E L P H I A I. That's what the platform is called. You can go out, you can chat with it. It's trained on every single episode of the podcast. You can ask it any question. It will go and pull from things that I've actually said to answer your question across over 350 episodes, that is, you know, 30 minute episodes each. So that's thousands of hours of audio. And you can ask it any question and in three seconds it will give you the response most relevant to your question. That's the power of retrieval, augmented generation. That's the technology that this is built on top of. But what that tool already enables is if you stray off the path of what it thinks it knows, it sends me an email. It says, I got a question that I couldn't answer. If you want to follow up with this person, and usually I don't because the question's something like, what's your favorite flavor of ice cream? And now it'll pull the Thomas Jefferson episode and it'll say, I'm not sure what Jason's is. Maybe try asking Mr. Daly directly. But Thomas Jefferson, I know what his favorite ice cream was. So we already have tech that can make the handoff to an advisor in certain situations. And that I think is probably the best balancing act from not wanting to let an AI assistant like go wild and just tell my client anything. I want it to ultimately pull me in if it's something that it couldn't find a very direct answer to. Because, like the, the context that we're going to give that AI assistant is all stuff that we have already said. Like, so just to take my, my sort of podcast AI assistant as a basic example, do I need to be afraid of it saying things that I wouldn't say? If all it is doing is providing clips that, that are things I've actually said, like, that's what these assistants are doing. They're sticking very close to your meeting transcripts, to authoritative documentation that you've given them. They're sticking very close to that stuff. But if you stray too far from that, it's gonna come and ask you for help. And if you think about the worst case scenario there, even if it can't answer the question, when I come into the platform so that I can answer the question, that AI assistant can already have gone out and fetched all the stuff that it thinks, I'm gonna need to be able to answer the question myself. So even in the worst case scenario, I'm ahead not having to go out and get all that information on my own. I love that one. The reality is even though like having AI assistance as part of what we provide to clients feels a little uncomfortable today, I would argue done right. If you stand up two professional services providers that are otherwise identical, the one that has an AI assistant that you can get immediate answers to for for some things that is a more compelling solution. It doesn't mean that you go to that for everything, but if there's a version of you that can be available 24 7, that's just valuable. It is. Okay, last last one I want to talk about because it's too good too good to skip if you're a software person. We'll put a link in the show notes to my software email list. We went through about about twice as many on the most recent email there Last the Narrative Daily Digest Imagine hopping into your PM or your ledger or whatever it is and having a digest of all the stuff that's happened since you last consumed the digest. Let's say that if I listen to this daily digest yesterday, then when I tune in today I want to see what's happened in the last 24 hours, right? If I last got into it a week ago, I want to hear everything that's happened in the last week. What is the 60 second version of that? What is the six minute version version of that? I want to know. Your client, they just signed their 8879 last night and that one is now waiting on Sharon to E file. Your colleague had a meeting with Jim yesterday afternoon and they resolved X, Y and Z. Outstanding requests. Tina accepted the proposal. That old invoice got paid yesterday. I want the rapid fire digest of all the stuff I want to know about that happened. And right now language models are more than good enough to produce this. You give it all the context, all the sources and it's very good at determining what are the spiciest things here. For the 60 second version, for the 6 minute version and it will pull that together into a narrative that would be like can't see TV every single day. Imagine the habitual product usage that would create where you roll out of bed or you rock up to your office. That would be the very first thing that you plugged into every single day. And I think platforms right now the closest thing they have to this is a dashboard. But a dashboard is like by definition rigid. It's not surfacing all of these things. Like it's not a narrative of this stuff that's happened in the last X timeframe this is even juicier if the platform has a mobile app and I can get this like over audio. Like give me an audio digest so that as I'm making my oats in the morning, I could pop this baby on the AirPod and get the full digest. Like you are going to have me returning to your tool every single morning just to get that pumped into my veins. That would be so valuable. And the tech is there. Like we've got it. There's no reason it can't be done today. Can you tell that I'm very excited about what the future holds? Almost all of these things are vertical specific things for accountants. It's why when I make recommendations, I'm recommending tools inside of our ecosystem. Because the ceiling of what general tools can do for us, it is much lower than what accounting like, like industry specific tools can do for us. And if in the future all we have are general tools like everybody else has, that is a bad space. Like that's a bad place to find the profession in. If, if the tools that we can use are the same as everybody else's tools, we need tools for professionals that help us serve clients in, in greater volumes that are solving very nuanced, like specific to how we work, specific to accounting tax very nuanced problems for us. It's an exciting time, man. It's honestly, this should only be scary for people who are like pretending this isn't a thing, who are hiding under a rock, who are unwilling to learn any amount of AI to see how it can help you. I mean there's already cool stuff like in the tax pros Discord, they were chatting the other day about using Perplexity for tax research. In my opinion right now Perplexity is the best tax research tool in the game, period. Even though it's not even a tax research tool. It's like the best AI search engine out there. There may be very, very specific situations. Let's link Perplexity down in the show. Not Perplexity, AI. There may be very, very specific situations where you have to do something highly technical into a specific source where maybe Perplexity can't get there. But now Perplexity, you can point it to specific sources. It can use the O1 model to do like multi level reasoning. It will draft a memo for you, even pulling in images in like less than 10 seconds. It's so, so good doing so many things that research tools just cannot do. Research tools today are still like text based, text based search matching. Like it's just matching the words that you typed in its database of documents, which is such an unintelligent way to do search. And it requires having to have the perfect vernacular and all this stuff, right? So many examples of that where AI are just vastly, vastly improved ways to get the same work done. It's stuff that we ought to be excited about. The only people really that are hard done by here are the ones that are like actively hiding from it either due to ignorance and, and I don't, I don't mean ignorance from a, I don't like an insulting standpoint, but they're just, they're clinging on to outdated information that we need to kind of get past and learn, push through some of the, some of the spookies. But then you've got other people that are also just unwilling to change. And those folks are, are probably going to have a hard time over the next five years. But the rest of us, boy, I have a hard time seeing there being, being any, any excess of accounting talent out there. Like, there's plenty of demand for what we're doing. Anything you've seen shipped in products, that's like super exciting. Drop a comment down below. I always try to highlight the company shipping the coolest new AI stuff. It's a great signal that they're investing in the future. That's it for today. Thanks for coming and hanging. Sam.