066 Someone Just Bolted ChatGPT To An Accounting Ledger

Somebody did it. Those maniacs. They bolted chatgpt to an accounting ledger. Oh, buddy, is this a good idea? Is this a bad idea? What does it mean for my 12 column paper? Come on in, the water's warm. Let's talk about it. Hey, quick programming note the day you hear this Tuesday, if you are at scaling, there's gonna be a meet and greet from 11 to 1 at the digits booth. Well, little ol me, I feel super bad every time for conferences. People are like, oh, I'm gonna be there, let's connect. Let me get you coffee or something. And I'm like, yeah, there's a lot of people. And then inevitably I don't connect with like 90% of people. I'm trying to have like a dedicated time and place at every conference so I can for sure be like, if you come here, I will definitely be there. We can glad hand back slap and knickknack paddy wax. So if you are there, please come hang. I'm recording this on a Sunday. I may not take the recording gear to scaling. I don't know. It's a quick trip if we have a couple days off after this. That is why thinking about it. Thinking about breaking the streak. 66 episodes. Okay. Uninterrupted. I'm doing my part. What does that even mean? I'm doing my part. There's so many of you that are turning in and hanging out at like tuning in and hanging out every day. It still blows me away. Okay, chatgpt for your accounting ledger. What is the world coming to if you missed it? I'll cruise through kind of the news release digits. I'll put a link to this overview I did in the show notes as well with pictures and stuff like that. So digits launched a chat assistant that connects to your accounting ledger. Specifically right now, QuickBooks sucks to be you0 firm. And this is the most high profile run anybody has made at chat for accounting. In the current paradigm of what chat means, chat's not necessarily a new thing, but GPT and the quality of chat has changed a lot in the last 18 months and to kind of like, I don't know, frame this for. Cause I think this came out a lot of people were like, cool, whatever. Why this is such a big deal to me. This was like one of my top kind of predictions that I did at the beginning of the year of the biggest stuff that we're gonna get from generative email to like the race to bolt chat to an accounting ledger. This is a big deal to me because ultimately if Somebody nails this like this may be the future of the accounting ledger. And we've talked in the past about how a lot of folks like Bill Gates has talked about this quite a bit where they think the, the user interface that we're kind of accustomed to now is kind of going to be a thing of the past. Because when the computer can just understand what you say or what you type, then you don't need all the buttons and you don't need the nested menus of all these different things. So we kind of are at like peak user interface and now starting to come down the other side where applications now will pro will possibly be chat first, be voice first and we'll never really go back to you know, like peak GM where the fancier the car was the more you had like 95 buttons on the interior of the car all with a little bit of chrome on the bottom of them. Just the most disgusting automobile design ever. And I think we're like coming back from that with like touchscreen stuff now. And I get there's pros and cons and you still want your BlackBerry with the full keyboard on it. Like I get all that. But stuff changes. So like the significance of this to me is like what does the next generation of accounting ledger look like? We've got Xero and QuickBooks as kind of like the Anchor SMB accounting ledgers. They've been around quite a while now like fairly early cloud companies, early ish. And I think we're all wondering like what does the next generation of tool look like? Very possibly something completely different that is not even like the traditional ledger that we're used to. And there's a Paul Graham quote that is, you know, you've already seen the future. You just already dismissed it as a toy. So that next accounting ledger will look like something that initially you're like pusha that does not do all the things that I need it to do for me until one day it does. And it's like better than the stuff that you're using, right? So if chat is truly going to take over the world and UIs are going to go away, then I am hyper curious about what in the world a like non UI accounting experience looks like. So like for like that is one reason why I think this is a big deal. The second reason is you've got digits who is like this very wild card software company in the accounting space. They've raised almost $100 million. They've done a whole bunch of different things. They like everything. Digits is very Razzle dazzle, like very impressive looking but I would say they're still yet to like find that really cool. Like anchor single kind of killer use case product market fit. They've got things, they've got a number of things that like on their own are very cool. Like we've talked about like vectors, vector search and embeddings and how that's going to change how accounting ledgers work. Like Digits is doing this right now in a meaningful way in the form of a quality assurance product where it will go out and identify outliers. Like stuff that was categorized wrong, stuff that was even coded to the wrong contact. They're doing a whole bunch of very technical stuff in the background of Digits is it was a whole bunch of Google machine learning AI nerds that went out and said let's figure out something killer to build for advisors. And you know what, that's a pretty darn cool premise on which to start a company as far as I'm concerned. And yes, maybe something like people have dirty windows, we need to clean their windows, like feels more like tangible. But ultimately if we as accountants are afraid of AI replacing us as accountants and I don't think most of us are, but like there's kind of that existential threat right Then we want the AI nerds on our side. Like we should be their biggest cheerleaders. So like I for one, like I'm excited that Digits has invested in this and is like kind of pioneering what new age chat for the accounting ledger looks like. Even though this is like absolutely a version one and we're going to talk about that, like what it does, but all the stuff that it doesn't do and like where this could be leading us. So to look at exactly what it is that they shipped. You can hop out to digits.com right now if you don't have an account, if you're an accounting firm, you can sign up as an accounting firm, put five clients on it for free, like not even pay a dime and you can take all this stuff for a spin. So it is a. This chat assistant now exists like in this right hand panel within Digits. And if you haven't used digits in the last six months, it probably looks wildly different to the last time that you used it. But built within the Digits product they've got like reports, they've got this kind of AI based quality check thing, a number of different things they're doing. But over on the right hand side now is this panel for this chat assistant and you can chuck any old question in there and it'll give it the college try or it'll tell you, sorry bud, that's not a thing that I can do right now. All you do, you connect it to your QuickBooks file. I think it may take, it used to take like a half day or something like that for it to like build the database and all that. Not sure if that's still the case, but that's all you got to do is connect it to your QuickBooks file. What it will do right now is it, it looks like it just has access to chart of account level data. So like account balances by period, the transactional detail. So like the full ledger detail for each account and customer and vendor level data. So for example, I can ask it to summarize like office expense over the last five months or show me all of the ledger detail for office expense from two months ago. But if there's like $1,000 transaction that I see there and I say what is this today? It doesn't pull that transactional level detail. It won't give me that. But everything that you would want to see maybe from a reporting standpoint and an analysis standpoint I think seems to be there. You can also ask it for like common like financial metrics, like current ratio, stuff like that. And it generally will understand that. It even does a little interesting thing where it responds and says, oh, that's like so if I say what was our current ratio last month? It responds and says, oh that's current assets divided by current liabilities. Current ratio for last month was blah, blah, blah. So it even like kind of gives you a bit of confirmation of here's how I'm calculating it. Another interesting thing that's happening here is like this and this will be a bigger, like a more long term discussion that will be interesting to follow. It is built on, I don't know if it's built on GPT or on Palm, Google's language model, but it does not have general, it does not have access to the model's general knowledge. So like when you go out to ChatGPT and you're talking with it, that is just all of the general knowledge that is built into that language model. But as we've discussed, you can build chatbots with the language model on your own custom context, your custom information, like where it's only restricted to this hundred pages of information that you gave it, it will still form all those things and talk like a human, but when it comes to like answering questions and that sort of thing, it will just stick to the context you give it. And that's the approach that they've taken here. So it doesn't pull from any general knowledge at all, but they have layered in some general like a general accounting knowledge base and I don't know where they pulled it from but like a user can do things like ask what is a balance sheet, what is an income statement? Like how do these things work? And it will answer these questions. And so kind of interesting like in a world of language models and everyone kind of using this same language models behind the scenes, you still have opportunities to build in proprietary context that you create. And this is one such example. We've talked about how even within our firms we can start developing proprietary context so that if we have a chatbot we surface to a client like the way it would answer questions would just be based on your expertise that you've captured. So interesting decision here. There are pros and cons. We're going to talk more about to allowing the general knowledge in versus just building it on top of your own custom context which is inherently going to be more restrictive than like all of the billions of parameters that the entire model is trained upon. Right? This episode is sponsored in part by Client Hub. Hey, Tales from the Hub is back. Circumstances Season 2 Episode 1 Buckle up. Recently on Tales from the Hub, you'll remember super smart accounting firm. Well, they adopted Client Hub to manage their work and collaborate with clients. Right after busy season they sent out a client feedback survey. They are super smart after all. And the results are in. Of course. Clients love working with Client Hub. They're eating it up. When is the last time you heard a client clamoring for a client portal? Right, the clients love the Client tasks feature. Super efficient and intuitive. No more emails flying back and forth. Clients love to have access to the super smart firm staff via messaging and Client Hub. Clients are saying the firm is a modern tech forward and client service oriented firm. You already knew you were, but now the survey proves it. And Client Hub playing a big role. Great job Client Hub. That's it for this week on Tales from the Hub. Learn more over at ClientHub app or click the link in the show notes. This episode is sponsored in part by the fine folks at Cloud Accountant Staffing. Do you hire accountants? Bless your little heart. Not the best part of the job in my opinion. Not something I ever enjoyed. Well listen, you can build your accounting dream team with talented offshore accountants in the Philippines that work 100% full time for your firm. Their accountants aren't freelancing or Contracting for multiple firms. They're all yours. They work exclusively for you and are incentivized to stay with you and your team long term. They're not going to get swiped. Cloud account staffing is 100% dedicated to the accounting industry and founded by a former accounting firm owner that understands your business, knows your pain points. They had to hire some accountants and they said you know what, we're going to build our own pipeline in the Philippines, going to pull in some super talented people and then open that up to other firms firms basically that's the story I've been talking about a lot about staffing, building more resilient staffing pipelines for your firms. I had staff in the Philippines like totally red pilled me to like oh geez, like we need to globalize the way that we get our work done. Check these folks out. Link in the show description Cloud accountantstaffing.com now what it doesn't do, like I said, it won't fetch you all the details for a specific transaction. And, and it is not like this infinitely flexible playground in the same way that ChatGPT is. So like the novelty of ChatGPT is the fact that you can ask it to do just about anything and it's gonna do its darndest to try to do it whether it actually can or not. But even stuff like take all these numbers and put em in a table and format them this specific way and then roll it up this way and that way and concatenate columns A and B and give me a table and it'll do all that stuff amazingly well. This will not do that. It is limited down to kind of more specific functions that it, that it will and will not do. And I think that the reason for that is in order to do this really well, there's a couple hard things that it has to do that like language models cannot do. One is the maths. It has to be able to infer what exactly is the user trying to calculate so that I can ensure that I give you a correct result back. And that's hard if you're especially the more open and more free form that is to be able to just like see into whatever that freeform response it got from the model was and like do math perfectly on that. That's very difficult. But second is the security component. So, and this is an interesting discussion because like this ultimately is a great model for how AI can be done in a very responsible, very secure way going forward. And right now the like the ways that like your carbons and Canopy, stuff like that. They've been doing it well, but like, this is another step in complexity. So from the press release, here's what I gathered about how they are approaching security, that like, they are doing everything they possibly can here. So prompts are not being trained into the model. That's an easy one and should be obvious. You chuck a prompt in there, it's not being retained in the model at all. And honestly, we are still grappling with whether that is enough, right? Like that might be totally fine. There's a ton of APIs out there that, that services we use make calls out to all the time. And like those API calls aren't retained. Like, so even that may be enough, but they're going further, which is good. Second is wherever possible, prompts will be anonymized. And we've talked about this anonymized information. Like there's, I don't know, a million different definitions of, well, what is, what is secure info and all that stuff. And, and you'll also get people kind of like making up their own personal versions of what that means to them, which is totally fine. At the end of the day, it ultimately has to be up to you and like your firm policy. But under hipaa, for example, if you got patient medical data and you take any identifying information out of that patient medical data, it's no longer sensitive information. And this is where you get the stories about like, oh, your medical information is being sold to Google and all these things. And it absolutely is. Like, that is what they do in their. Unfortunately, under today's rules, completely within their rights to like strip your information off of it and then go sell it to a third party. That is anonymized information. Under gdpr, anonymized information is not sensitive information. Under California's privacy rules, if that information can't be died back to a specific person, you sidestep like the requirements of how you have to manage that information and what you're allowed to do with it. So an anonymized prompt under today's rules, not saying maybe the rules don't need to be different. Generally anonymized information you are able to do anything with. Third, there are some situations where a prompt can't be fully anonymized or you need to recover some of the details from that. So for example, let's say you asked it a question about five customers. Well, you don't really want it to send the prompt to the model with all those customer names, but when you get the response back, you want to know which customers it's referring to. Right. And this is an alternative to anonymization called pseudonymous pseudonymization. Close enough. And the way digits frame this is they obfuscated was the word they used. Identifying information. But pseudonymization, gosh, is like a well defined approach under GDPR where you basically substitute the sensitive bits. You send that off, but when you get it back, only you have the lookup table to then be able to put the identifying information back in. So for example, a simple prompt, Bob has four cats. But we don't want Bob and we don't want cats going to the model because we don't want anybody to know that Bob has four cats. So we substitute the words BOB and the words cats with some sort of gibberish. I do that on my servers so that then the prompt is like gibberish has four gibberish. I send that prompt off to a model, it didn't see Bob, it didn't see cats. It sends it back with the two bits of gibberish intact. I replace that with the BOB and the cats, and now I've got the response without having ever sent the original like sensitive stuff to the model. So they're using a combination of that and anonymization to ensure that like nothing identifying goes to the model, even though the model isn't retaining any of those prompts. And I think this, it'll be interesting to talk with somebody smarter than myself about this, but I think this may ultimately make like the wild west version of an accounting ledger that's more akin to jack GPT where it can do absolutely anything. I think this might make that hard because ultimately how will a computer be able to figure out like what is identifying information and what is not? For example, like one of the things digits has in their transaction review product is if it identifies pii, like a Social Security number or an ein or something like that and a memo line in an accounting file, it will flag that and say you probably want to remove this. And that makes sense for certain things like a social that are a kind of fixed structure. Computer can pick that up. But when it comes to any type of identifying information, that's a hard thing to just broad brushstroke have a computer remove. Right? And I think that's maybe part of the reason why these initial passes at chat with your ledger may be a little more restrictive and not totally free form. And that's kind of a bummer. But again, this version one, and as with all things AI like, and this is what's so exciting about this. This is the worst it will ever be. It will never be worse than this. And this I think already has value. Like, they're doing some cool stuff here. Another sort of novel thing they did is they're actually like embedding these graphics and kind of dynamic tables within the text response. So it's not just a vanilla text response like you get with ChatGPT, it's like a helpful visualization that is paired with each of them. So they're introducing some cool new concepts here that I think they can continue to build on. Specifically a few things. I know they're working on mobile support. So being able to chat on mobile, follow up questions, the ability to return reports like balance sheets and P Ls and then class and location support as well. But the main idea is like, get this in the hands of as many people as possible and then tell them, like, what is the best version of this for you? Like, what is the situation where this would actually be helpful for you? Because accountants and like the. The useful version for a business owner and an accountant I think are very different. But for accountants, we already know how to get the information that we want generally right. Like, we understand the accounting software in and out. The only way we're ever going to use something else is if it will either in a quicker or more meaningfully meaningful way, pull that information together for us. But I would argue in the same way that nobody, the day before ChatGPT came out, was sitting there going, boy, I show up, a chat assistant comes out tomorrow that can write my emails in a pirate voice in the same way that nobody was clamoring for that thing to make their life easier. I think it is admittedly hard for us to be like, oh, finally an accounting ledger chatbot that will do X. So like, this is gonna take some time. The reality is like, to think through, like, what are the useful ways that this can all come together? And then long term, like, if we're going full, like big brain AI theorist, how does this ultimately even replace a ledger product? Got a lot of steps to go before that, but I thought it'd be. It was like, interesting. I put this stuff all out there and I really wanted to like engage with accountants on, like, what are your thoughts on this? Like, is this weird? Like, can you think of use cases for it? And a couple common themes. And this always comes up whenever I talk with chat on the ledger. We've seen this before and it is true, like, chat bots were like the hotness five years ago. Facebook was really pushing them. And it was like, oh, this is the end of the web browser. And it was like, everything's going to get rolled up in chat bots. Until it didn't. And so you had QuickBooks with their own first party chat assistant. Even a voice assistant. I don't know if that's still a thing you can actually use, but it was around like for a minute. Chat AI, I think they've pivoted to something else now. But like initially the premise of that was natural language processing for your ledger so that like client could ask a question like, show me the office expense detail for last month. And they didn't have to understand how to work the software. This episode is sponsored in part by Firm360, the practice management tool that gives you a 360 degree view of your practice. Talked about this in the past. Project management, documents, time and billing, all that stuff's in there. Manage all that stuff in a single place. Story time, though. Let me tell you about Lee Lee's team. They were looking for a solution to modernize their firm's processes. They'd grown 30% year over year for the past five years. Holy mama. And their processes were no longer able to keep up with the amount of staff and projects they had going on. Real talk your systems. Like, yeah, no, that changes with the size of your firm. They still had some of their documents and paper copies and they were going around the office to hand off projects to team members. Yuck. Once they were on firm360, they were able to get all their digital, all their, all their, all the goods, all the stuff in a digital format and save a ton of time because they had it all in a single system. Got visibility into project process. Everybody's on the same page about what they're supposed to be working on. They even implemented the secure client portal, allowing them to deliver documents and collect payments digitally. You know what took them? Two months. Two months. Not bad. Not switching PMs, not fun. But knocking that out in two months, that's pretty darn good. Hey, let's be a little more likely, huh? You hustle on spreadsheets to manage that stuff. Knock it off. Shuffling papers around the office. Modernize your practice. Learn more about this1@myfirm360.com Now I made a joke where I posted this on LinkedIn and in my newsletter that was like, here's what I said. Astute observers of tech trends are quick to point out people were bolting natural language processing to ledgers five years ago. And it's True, you also got a staph infection from a VR thing in a strip mall when you were 14. But that doesn't make the Apple Vision Pro any less interesting. Like, honestly, I'm not that concerned with past chat implementations of Ledgers as, like, precedent for where we are today, because put your finger on another chat GPT that has, like, ever, you know, had this sort of impact. So the capabilities of what we can do with chat right now are just. They're fundamentally different. We're seeing the plugin ecosystem within chat CBD absolutely exploding. It's over 500 plugins now. When we did that plugins episode, there were maybe 30. I think that was less than a month ago. There were not. Maybe there's less than a hundred, I don't know. But, like, the number of plugins coming out for chat GPT is ridiculous. Right now, I am finding myself using plugins more and more mainly to chat with documents, to chat with videos. I will actually now put the URL of a podcast of mine in there and be like, did I talk about this thing in that episode? Because it's only a matter of time before I do an episode on something that I already did an episode on and just forget. But chat today is not what chat was even three, even three years ago. So, like, I am very interested in exploring what does chat look like integrated into the stuff that we do, into accounting stuff. Right. But ultimately, and I think it was Nicole McKenzie that said this show me how to make money with it. And I do. Like, that is like the very practical accountant, I think, like, how's it going to make my life easier today? Sort of response like, and I get. That's how, like, almost all of us are wired. Like, you've got. You're like, 20% get off my lawn. 40%, I'm too busy to deal with this right now. 20% get the intern to look at it. And then 20%, hang on, I might actually be able to make something funny that I could send out in an email chain with it. And today, like, no, I don't. Like, I didn't see anything that would, like, make my life easier. What? I don't know. And honestly, maybe the more interesting thing to me right now is are there any of my clients whose lives would be made easier by it? And like, to put this into context, when I initially did that, like, AI prediction roundup at the beginning of the year. Yeah. On January 13th, I did some polls just around the idea of, like, what if there was a chat assistant that could just be the buffer between you and a client. And all of the questions that we get to give a client a place to go, ask the questions that either they ask you that are very easy, like low hanging for questions, but equally importantly, the ones that maybe they're afraid to ask that are like the dumb questions where they don't want to get a bill for it or something like that. Right. So I put this question out there and I said it was in January. I said, if there was a ChatGPT experience for an accounting ledger that you could flip a switch on and turn on tomorrow for your client, would you do it? Would you turn it on? 40% of people said absolutely. 45% said maybe for some. And only 14% of people said nah. So I might. So, like, I think our initial view of this is like, how's this gonna make my life easier? And I would argue it may not have to make your life easier if it makes your client's life easier and maybe streamlines your client relationship a little bit, there's value there. So TBD to me, like, whether, like, oh, is this a thing that's helpful for my business client today? But then the other interesting thing here that I don't think anybody has hit on, this will be interesting to follow as like chat tools get more nuanced. And it's not just chat for accounting, it's chat for this specific type of person, maybe in this niche. And, and you know, on the advisor side or on the small business side, what I'm interested in is the like advisor in the loop version of chat. So you're going to have situations where the language model will be perfectly confident answering a question for a client. Show me last month's office expense detail. That's super easy. It can go do that. But what about when the client asks something maybe a little bit more complex where the language model isn't quite sure of itself? What I would love to have is that chat interface where the client can come with questions, but anything beyond a certain complexity, the advisor gets pulled in. And I would love control over where I set the threshold for that complexity, whether it's I need to be pinged every single time or whether it's Mad Max, client can ask it anything and it will always immediately respond. I would love to be able to set the threshold at which I need to be pulled in. And here's why I think this is smart. And what I really like about it is worst case scenario, like total fail state, the client asks the chatbot a question and it won't Give it a response it doesn't know, so it chucks it over to you. But when I hop in to answer the question, that AI could have already pulled everything that it thinks I need with which to answer the question and maybe even suggest a few responses. So worst case scenario, I'm still well ahead of the current state of things where I get emailed that question and I have to go then log into the ledger and answer that form for them. So I love the idea of like, what is an Advisor chat service look like? And one of the common talking points here is. Is like, well, what. Why don't we just wait till QuickBooks rolls out their version or till zero rolls out their version and I don't know, like, within those products that they would roll out the Advisor version of chat as much as they would roll out, like, the version for small businesses. And technically, I would say that's what Digits has done here too. Like, they say they're a platform for advisors, but right now the chatbot went live in the small business product, and in the Advisor product, they're not any different. So ultimately, like, is there space here for somebody to make. Hey. And somebody to make a cool product that's really helpful for us? I think absolutely, yes. Even if that chatbot isn't helpful to me, buddy, it's helpful to me if it's an assistant that can answer my client's questions. Because if they're rolling out to ChatGPT right now that can't see into their QuickBooks detail, that can't see into their ledger history, it can't see into all of that. To be able to answer questions specific to their history is going to take a product that actually connects to their ledger, and that's something I can enable for them because I have access to that. It's never something that's going to be part of, like, the general knowledge of a language model. So to summarize, is it going to change your life right now? No. There may be some clients that it's like, groundbreaking for and really helpful. I was. I've kind of been in the camp where, like, I don't always give clients direct access to the accounting ledger. That can backfire. I think, in fact, I. And so, like, that's one lens through which I like reporting tools. That sort of thing is as I can kind of keep the client at arm's length while still giving them access to something. Digits could be a great version of that. I think that's kind of the version that they're going. They've got like an uncategorized transaction product now where they. I think they want eventually to like, get your clients onto this platform so that you can send them questions that way. But then ultimately, if they like, if the client's able to answer some of their own questions through the chat assistant, that becomes valuable to me. But they got the version one out there. I would encourage you to kick the tires on it. If nothing else, then to give them feedback and tell them this is useful for me, this is not useful for me, or even this is useful for my client, not useful for my client. We want AI for accountants, right? Like, and this is why I will have a really hard time being a negative Nelly about any of these AI releases. You know, we've been very complimentary of Canopy of Carbon, of all the folks investing here. We got people investing in this stuff. Like, let's support them and like, tell them exactly what we want. That's going to shorten the path to getting really cool, meaningful stuff in our hands, I think. So that's it for today. If you are at scaling, come hang out with me. We might be back tomorrow. We'll see if I take the gear to scaling. I don't know yet. Either way, I'll see you soon. Bye.