168 My 2024 AI Predictions for the Accounting Profession

The big day's here. We're gonna run through my 2024 predictions for AI for the accounting profession. Boy, was last year a big year. Here's what we're gonna do. We're gonna recap 2023's predictions. Check where we're at on kind of the roadmap on those things. A lot of them we got, some of them are still in progress. Next we're gonna touch on the technical, the more nerdy technical developments that are happening by the day right now. To then inform number three, how does this change the tools that we use and the ways that we work with clients? I am so excited. Come on in. Get the old, the old crystal ball out. Okay, I've got a, I got a fairlife protein chocolate milk here and a little cup of Nerds gummy clusters. A 12 year old inside me on full display. Sidebar. Nerds Gummy Clusters S tier candy. There may not be a better candy in the candy game right now than Nerds Gummy Clusters. If you know, you know. Okay, let's set the stage with 2023's predictions. Did this on January 1st of last year. It actually sets the stage for a lot of what's going to happen in 24. So let's run through these stuff, see how we did highlight some companies who are giving us this stuff because a lot of 24 is going to be a continuation of what we got in 23. My predictions last year, number one was generative email replies, basically using AI to generate or suggest like a draft reply to an email. So we were talking about this back in November of 22 as really being the biggest thing that could be an AI unlock for accountants. We got it. Canape, I think was the first practice management to ship it, to ship generative AI when you're drafting your emails. I believe Carbon was the first to shift generative replies, which is different. It is when it actually sends the whole like conversation chain to the language model to suggest a reply, not just to draft an email from scratch. No client hub's been working on this stuff. We've got a number of more folks getting in on the actual. But when accountants spend so much time on email, I still think this is maybe the biggest short term opportunity where accountants can save a whole ton of time and software vendors can put way better tools in our hand to like draft these replies. I mean, the examples I gave last year, you get the email, can you Send me my 2020 tax return? Give me the suggested reply. Hey, Mike, here's your 2020 return attached and then go pull it from the file system which is part of my practice management system. I even had a little tag on here. See you next Thursday. So these generative replies in our practice management system using AI to see into what's on my calendar, to see into past email conversations with the client. See the email my colleague sent the client yesterday. See the status outstanding status of projects. That assistant that helps you draft emails. We've got version one of it right now in a number of systems. But it has so far to come still. And I think we're going to see some big investment in that this year. Number two prediction last year was putting a chat assistant on top of the ledger. So Intuit announced this with Intuit Assist. I don't think they've actually shipped it yet. They had a demo at their conference last November. But Digits shipped a pretty solid version of this three or four months ago. Now has it taken the world by storm? I don't know. But I think we definitely learned a lot by them shipping that. We learned that a chat assistant for an accounting ledger that is for a small business owner is very different than a helpful assistant for an accountant because accountants generally know all the buttons to push and where to get the answers. We already know all that stuff. So an assistant that would be helpful for us is actually completely different than an assistant that is helpful for a small business owner. And right now Digits is chat assistant. It's really probably more oriented toward the small business owner trying to make sense of of their books. That being said, they did some interesting things from, you know, chat GPT is just text back and forth. One thing Digits does is it actually embeds little charts like kind of mini tables inside of those responses. That's an interesting way to make those make financial stuff more understandable than plain text. I think the notion of doing accounting work in like a chat dialogue or a like voice speech back and forth with your computer seems pretty abstract, right? Because a lot of numbers and needing to see reports and all that. But digits embedding these kind of mini reports and charts into a chat assistance interactions, that was really interesting. So we got that this last year. The most disruptive version I see of a ledger chat assistant in the short term is probably a system that does that really well for the layperson for your small business owner. We'll get to this on my predictions, but I think we will have a lot more AI advisors in 2024 that are services that you can connect to your accounting ledger or upload past tax returns to that are pretty good. Not as good as a pro that's going to dissect your situation. But right now, what do people have access to? Like, if you don't have an advisor who is going to advise you on the stuff that's in your accounting file or the stuff that is in your tax return is really hard. I mean, you really can't get any sort of even basic answers. What are you going to go like comb through tax return instructions or just pull up your dashboard in your QuickBooks file? Like that is not good information oftentimes because the underlying bookkeeping isn't correct. And so the most interesting or disruptive version of a chat assistant for your accounting ledger in the short term is probably a consumer facing one that's much more helpful and will answer a lot of basic questions for business owners. Number three, prediction was getting a big step closer to fully automated bookkeeping. We did a back in June from a from a hotel bed in Las Vegas, I ran through how embeddings work, how vector search works. I think the episode's called something like the path to fully automated Bookkeeping. This is a technology embeddings and embedding models and vector databases are a technology that's every bit as big and impactful to what we do as language models. Basically makes it incredibly fast to semantic search. Huge volumes of information that is like how semantically similar are two different things which comes up in research. And in all these different things, one of them will be it massively improves the ability of our accounting ledgers to automatically classify things transactions. And what this will ultimately mean is a move, a move away from rule based classifications and accounting ledgers. So today the only way to really automate your classifications is to set up and manage a bunch of bank rules. But when you have to do that across 10 clients, 100 clients, a thousand clients, there's no way to keep those rules in sync across a bunch of different files. Or when you make a one an update in one for that to populate across all of your files. What this change will ultimately mean is a move away from bank rules to classifications that happen automatically just based on what you classified in the past for that client and bigger picture like within your firm or you as a user. So more of an ability to correctly guess how it should be classified next time. One of the big really impressive things with language models now and a number of new AI developments is what's called, you know, few shot or zero shot learning. That is you give something just a few examples and then it can Continue doing that thing really, really well. Whereas in the past with AI and machine learning, you had to give it huge volumes of information to train it on what to do. The improvements in few shot and zero shot learning now mean that if I classified, you know, something from this merchant this way a couple times in the past, do you really need me to make a bank rule for that? Or can you infer that? Especially if you can see into the other similar accounting files I have in my in my firm. If you can see the other 20,000 classifications I've ever made in the platform, just think about how the state of how dumb that is right now. What do you think the total number of classifications you have made in an accounting ledger are and how many of those have been repeat merchants with repeat chart of accounts? The same things over and over again. This is a different level of intelligence that will just make that classification so much better and not require you to set up rules that will learn from the way that you actually code things. Do we have it yet? No. QuickBooks and Xero honestly have been saying for almost a decade now they're going to build a meaningful version of this into their platform. What changed in 2023 actually a month after I wrote these predictions is embeddings got really, really good and really cheap. Like 100x plus order of magnitude cheaper to do this. That's just going to make bookkeeping fundamentally better and more streamlined. Prediction number four Generative advisory prompts. Be able to to craft suggestions in the context of doing advising. I don't know that we have this in any of our reporting tools yet. Is anybody like generating insights? That's not true. Digits actually does this digits has a reporting platform and for like the executive summary they have kind of some generative things that they put in there. I could see this in your tools like you know, Fathom and Reach Reporting and Spotlight and stuff like that where they help you draft your executive summary to the client by like suggesting these prompts based on the underlying data. I haven't seen a lot of that in a while. It wouldn't surprise me if they did it. The bigger thing in that space is probably just the improved quality of like advisors that anyone will be able to get access to. AI advisors. It has you connect your accounting ledger and then you can start asking it questions. That's probably going to be the bigger thing in that space. Prediction number five this last year was tax research would get turned on its head. Boy, all of the tools are there for this to be so, so much better than it is and it's not happening yet. One of my biggest disappointments in my, my year end review of the best new stuff we got this last year, this was like one of my three biggest flops was tax research. Vendors have just not done anything good with this yet. Prediction number six, improved document extraction. That's absolutely happened. Like OCR is basically solved now. We got the GPT4 Vision model from OpenAI a couple of months ago. It can read pretty much any handwriting. It's phenomenal. Especially in cases on tax documents where you have the same numbers repeated several times to like, like validate that. Yes, the way I read that number the first time was created or was correct. This is such a huge step beyond OCR and what we've had. The whole like the double checking that has always been required when there's an invoice or a tax document or something like that. That whole workflow of double checking what gets extracted. I think we may be getting really close to, to moving past needing that double check. Prediction number seven was AI as a feature. Most software companies, I said throughout their roadmaps to pivot to developing AI as a feature within their app, that's absolutely been the case. Like everybody is bolting on AI as a feature now. It's funny, 12 months ago a lot of companies were still saying, nah, we'll wait and see. Not that big of a deal. Very few companies now that have not like genuinely pivoted a huge percentage of their work to AI development. So that's enough of last year's stuff. What are the cool technical things happening right now that are going to lead to some meaningful change to the tools that we use this year? Here's my predictions on the more technical development front. I think we'll probably get GPT5 late in the year. Building GPT5 is a massive undertaking in a lot of different ways from the amount of money and energy and a lot of different things it's going to take. But more interesting might be GPT4.5. In fact, if GPT4.5 is a reasonably big step, frankly I don't know that they need to rush GPT5 this year. So Google Bard. Not Google Bard, Google. What's Google's big new model that they've been teasing that just came out? Gemini. So Google drops Gemini and it has a few different flavors and the only flavor they released is the equivalent to the old GPT3.5. And the version that's supposed to be on par with GPT4 isn't even available till next year. And they're not even going to give it to everybody. And now people are saying, like, no, it's not actually going to be as good as GPT4. You know, what is wild? GPT4 came out last March, it's been almost a year, and nobody is even close to the quality of what OpenAI released back in March. And so part of this is like, what does OpenAI even need to do? Like, do they even need to, like, keep pushing the envelope further? But I think we'll either get 4, 5 or 5 this year. 4, 5 is more likely because I think there's more ways they can augment what they already have with four in a way that is really useful. Going to talk about that in a bit, but we will get one or the other, I think, and it's going to be another pretty big tipping point where it will augment the ways people work in a really impactful way. It will kick off another big round of development. And the other tools that we use, because all the software vendors, not all of them, a lot of the software vendors that we use, like, they are all working as quickly as they can to adopt these AI changes into their apps. And this is like, this next version of GPT likely unlocks a whole nother level of capability and different ways that it can be built into our apps. So that next model, if you think of everything, GPT 3.5 changed everything, GPT 4 changed, going to be a big change and maybe where we pull an assistant in to help us with our work. I think the, the conversation around job displacement probably gets a lot more real with that next model. This is, you know, the, the best analogy to me that I've heard is along the lines of the Industrial Revolution, where jobs are going to change a lot over the next decade. And I don't know that it ultimately means that there aren't enough jobs for everyone. It just means that the types of things that we use humans for will change. And on a micro level, there's going to be some really hard aspects of that. But like the Industrial Revolution, I don't know that anybody looks back and says that wasn't worth doing. Even though on the micro level there was a lot of suffering in the Industrial Revolution. That next model we get from OpenAI, I think kicks off that discussion in a more meaningful way. Next prediction from me, honestly, a growing market for uncensored models. The models we're using are getting more and more locked down by the day. Claude has gotten super locked down in the last few months. I just heard A story of somebody who was going through a big writing project with Claude and they had done a ton of work on it. And then Claude says, I'm no longer going to help you with this. I've decided that it is against my moral ethics to support you in this writing project. And it was like some like, story that they were writing. And it literally just said, I'm not okay helping you with this project anymore. And peaced out. That was it. And so then you have like this human arguing with an AI, saying like, no, this isn't wrong, like, you can come help me with this thing. Romeo shared the other day a screenshot of ChatGPT telling him to go Google it. So that like he asked a question and its reply was telling him like, why don't you just go Google it? Which is hilarious to me. And obviously, like, I'm sure that's in the training data in a ton of places when, when it picks up the Reddit comments and stuff like that to say, just go Google it. But obviously open AI is, you know, a lot of people complain about how their models have gotten more restrictive and it says, I'm just a large language model, I can't do this. And that the more that happens, the more these companies get big and have more to lose by language models saying things they don't want them to say, the more of a market there will be for other things, not, not like uncensored models from the standpoint that you can do bad things with them, but just because, like there's going to be certain tasks where you just beat your head up against the wall because for some reason it won't help you do that. And so we're going to keep seeing more people in the space and more of an appreciation for models that are not restrictive. You know, people often ask me, jason, who is this episode sponsored in part by today? This episode is sponsored in part by liveflow. Did you hear the news? Lifflow just launched a consolidation product. You actually might have seen it on the main channel recently. We did a whole demo day of it. Liveflow's automated multi entity consolidations. It's beyond simple to use. You can easily map multiple unmatching chart of accounts from multiple QuickBooks Online companies into one standardized report. And once it's set up, liveflow's gonna get to work updating linkins to consolidations automatically in real time, the realest of time, so you can focus on analysis using instantly updating data across entities. Lifelow can even consolidate financials that are in different currencies that sounds disgusting. Yikes. And it doesn't stop there. Liveflow offers flexible, powerful reporting tools. Create customized dashboards that meet your specific needs, you little snowflake. Build executive presentations, cash flow forecasts and more with just a few clicks. The consolidation thing is actually super cool. You haven't seen that yet, check it out on the main YouTube channel. And thanks to LifeLow for sponsoring the pod. Next prediction models are going to become more purpose driven. So as the ecosystem has been built to kind of support people incorporating GPT and all these things into their apps, it's now easier than ever to fine tune language models. You can actually fine tune your own GPT model on OpenAI's website without code. Like that's a thing you can go out and do now, which is really cool. The process of training models from scratch, fine tuning models, using things more technical, things like Lauras, to target the functionality of a model for a more specific task. That's all getting easier than ever before. Rather than having to figure out how to do all that stuff yourself, there's now a lot of off the shelf tools that will help you build your own version of a model for a super, super specific task. So maybe that is extracting information from a W2, maybe that is being an advisor that is connected to a general ledger. Maybe that is looking at a balance sheet and offering insights. Right now we have the general tools like your GPT4 that can do some version of this. But this year there will be much more investment in purpose driven versions of those where they do that job better because they're just like geared up to do that job. And that's particularly helpful in the tools that we use, like when you're using them in a more contextual way. So an example we talked about is like advisory insights in a reporting platform like an assistant that is optimized for identifying, you know, potential errors in a month on close or opportunities for conversation. If you are. If you do like VCFO work and you do advisory for your clients, using that assistant as a second set of eyes on the books to like potentially offer advisory insights that you could pick up and discuss with the client. Models that are more trained for that specific purpose are going to be more helpful next prediction, better fetching or retrieval incorporated more seamlessly into models. So we've talked about this quite a bit on the podcast, but ultimately language models are not knowledge bases. We don't use them to get truth from. The best way to get to concrete answers with a language model is to provide it with context like a piece of legislation, and then say, answer this question according to the context that I provided you. That is the biggest step that we got from GPT 3.5 to GPT 4 is GPT 4 will stick to a body of text and only answer questions according to that body of text. And that is a fundamentally better way of using a language model, because you can ask it for citations, you can tell it to just use direct quotes. And at that point, the language model is working with bodies of text the same way that humans do. So rather than a research assistant going out and pulling together all these quotes and paragraphs and all these things from authoritative sources, you give the authoritative sources to the language model and you say, you basically outline the exact research process that a human would do, pull out all the direct quotes that are most related to this question, list them all out, give me citations to where each of those quotes are in the authoritative documentation, and then using all that information, answer the question, maybe even build a counter argument for the answer that you gave me. Get all of that in 10 seconds. And that is all driven not by what the language model knows. It's driven by what are called retrievals or the information that it pulls from the context, or the documents you give it that it uses to answer the question. And so in something like what we do, which is grounded in, like, you know, hard technical rules, a language model on its own is not particularly helpful. We can use it for, like, drafting things and writing and all that. But that's, that's great, but that's not helpful for getting to what is the right answer along with the source for what the correct answer is. Right, this is, and this is where retrievals come in, is when it can fetch from documents, fetch from bodies of text that you give it what the right answer is. And so retrievals, I think, will will become more seamlessly integrated into like our chat experiences right now. Oftentimes when we're talking with the chat bot, like it's just the raw language model. It's not pulling in knowledge from other sources. Custom GPTs, OpenAI launched, in fact, we're getting, we might have the GPT store by the time this comes out. Anybody can go out and build a custom GPT without code that is a custom assistant to do whatever you instruct it to do. And the most interesting thing about that to me was that you can upload documents to it so it can look in the documents that you give it and use that information when you're having interactions with it. So you could build a Custom GPT for IRS publications and chuck all the publications in there and chat about all those publications. Now, there's still limitations around retrievals that we're navigating. Things like, well, if there's a table in there, it has a hard time like semantically understanding what's happening in a table. Or if there's a flowchart. You know, IRS publications have all sorts of little flowcharts and stuff like that. And if you just copy and pasted all the text out of that page and put it somewhere else, that flowchart wouldn't make a lot of sense. Right. So there's ways to mitigate this with things like vision models that you show the table or the flowchart to and then it semantically describes like each path through the flowchart or semantically makes a sentence for each line on a table. You can manage it with stuff like that. But retrievals being more integrated into how we work with AI tools fundamentally makes them more correct and more grounded and stuff that we can trust. A really tricky one here is tax returns. You know, if you think about the layout of a tax return, it's a bunch of boxes and like text in weird places. And like just the text of it. If you just grab that PDF and you copy, paste the text of that and put it somewhere else, it's meaningless. And this is where we probably have to go to like purpose built models to make sense of tax returns or potentially even vision models that know what they're looking at when they see a 1040 and can pull stuff off of it. To date, that stuff's been garbage. Like, it hasn't been helpful at all. But we now have the tech to do it. We're just catching up with building meaningful applications of that tech for the problems that we have last and the biggest. I've been railing on this since like the second episode of this podcast. Agents and language models having more agentic abilities. So agents are tools that work an app for you or work your desktop for you. And the way they work is literally with this like conversation with the language model where it says, the user says print balance sheet and a P and l from QuickBooks for me, cash basis for the last calendar year. That goes to the language model. And then the language model decides based on what is on the computer screen right now, what, what's the most logical next step. So it looks at everything that's there and it decides, okay, logical next step is to click the reporting button. Now it's got the data that's on the Reporting page. What's the logical next step from here? Well, here we've got a balance sheet report. I'm gonna click on that on the balance sheet page. Now what's the most logical next step here? It's genuinely reasoning through each of those pages the same way a human does, using the data that's on the page. So generally this will be scraping the HTML that's on the page and inferring what that next action ought to be and going through these step by step. I've shown off a few different agents this year. A few months ago we played around with one called hyperwrite and it could do just this. I could instruct it, it's a Chrome browser based agent. I could instruct it inside of QuickBooks Online, go print me a balance sheet and a P and L to a PDF for the last calendar year cash basis. And it could do that. When I kick off that task, it launches a Chrome window in the background and it took maybe two or three minutes for it to complete that task. And it did it correct like 75% of the time. And it's pretty inefficient, pretty costly I guess if you compare it to a human, maybe not, but it's pretty inefficient, not super fast and not like super optimized for that sort of thing. And so one thing a lot of folks have been talking about with OpenAI's next release of GPT is back in the day, back in like when you were back in GPT3 and DaVinci and before all this stuff became mainstream, the way these language models worked was actually like to complete something for you. So they were what was called completion models. So if I wrote half a sentence, this could presumably finish the sentence in a way that made sense. And so all the prompting that you did was basically setting it up to successfully continue this thing that you were writing. And then ChatGPT came around and they realized, oh, there's like a more interactive version of this, it's more helpful. So they, they modified the way that they trained models to be optimized for chat interactions rather than just completing a statement. And so that was a different type of model then was something that was optimized for having an interaction the same way a human user would and asking follow up questions and all, all that as opposed to just completing a bit of text. Now what a lot of people think we're looking at next is a version of these models that performs better as an agent. That is, it can be given a set of instructions and Some context, that might be the HTML on the page, that might be the help documents for how to use this app and more efficiently return what the next action ought to be. Because that's how these agents work. You give them a bunch of information and say, make the next decision. And that next decision could be clicking a thing on your screen, it could be typing something into your keyboard, it could be launching an app. And as soon as it's capable of reasoning through each of those things, we're talking about, like, human levels of, like, intelligence and being able to perform basic tasks. That's a hugely important threshold. Right? Right. So biggest prediction for this year is a ton of progress on the agent side. Like, that's kind of the holy grail. The best and most meaningful application of these language models is not a buddy I can talk to and, and interact with and be a helpful assistant. That's all great agents are literally picking up tasks and doing them themselves from end to end. So if you think about how you, you know, may assign a task in your practice management system to a fellow colleague, and maybe that's a recurring task, whatever it is, in the near future you can assign tasks to agents and they will do those things for you. Super interesting. And the whole ecosystem of what that looks like is going to be really weird because who builds the agent on top of QuickBooks? Probably not QuickBooks. Like the day that somebody builds an agent that is capable of reconciling the bank to a bank statement in QuickBooks, that's huge, right? Like, how many, how many times each month around the world does somebody reconcile a bank statement in QuickBooks Online? But the moment we reach the tipping point where with a Chrome extension, you can upload your bank statement to the Chrome extension and it can go, you know, from the QuickBooks file it's already in, like navigate to that and like, you know, annotate the PDF and take everything off in QuickBooks Online, we're really close to that. But there's some of these tipping points where as soon as that thing is good enough, that changes a lot of how we work. And it is maybe a little problematic for the ecosystem because who made that Chrome extension that can do that thing probably won't be into it. They probably won't be the first one to do that, you know, for our practice management systems. Will it be practice management systems building helpful agents that can go into other apps or go look at the QuickBooks API to pull a reporting figure into an email reply, suggest and stuff like that? Right. It'll be interesting to see who makes the most of the agent stuff. Okay, that is like my prediction for technical developments, the nerdy stuff. Now for the tools that we actually use, how will they be most impacted? Let's talk about the accounting ledger. Biggest thing will be moving away from that rules based classification methodology. The ledger being capable of classifying things just based on my past classifications. Maybe based on past classifications I made for a similar construction accountant in another file. It's all the stuff we've been asking for for the last decade made easier by specifically embeddings and vector search. Now the devil's going to be in the detail like how good this is. Devil's going to be in the detail of the implementation because like what is that confidence threshold where it should classify it versus having to ask me. And like there's, there's. I, I don't mean to make light of a ton of nuance that goes into building what is the right way to do that, but the biggest long term impact that has on the ledger is right now we have an entire generation of cloud ledgers that are built around classification. That is what they are. At their core is classifying transactions and reconciling the bank. What happens when that's table stakes? What is a ledger? When the ledger bit gets vastly simplified, I think these platforms just move to being more all in one sort of money spend management type of platforms. So reporting takes on greater importance. The quality of an assistant that you can talk with that will give you intelligent answers about the data. The platform's ability to automatically recognize errors and outliers and how things were classified. I mean how many, how many AI companies in the last five years have come about and then died where the whole premise was oh, just connect your ledger and we're going to give you insights. And the blocker to that is always the quality of the underlying data. So like the nuance in what platform is ensuring that the underlying data is correct like that it's, it's going to come down to that sort of thing now rather than who's giving me a reconciliation experience that I enjoy the most because I think that's going away second tool that will be impacted practice management systems. Honestly probably for pretty much all the reasons that we outlined last year. The biggest low hanging fruit here for me, and I don't want to say the low hanging fruit to imply that it is a simple thing to build email, client communications, generative reply drafts, even the assistant that will have a helpful little suggestion alongside me drafting that email to say hey, just so you know, here's a summary of the email that Tony sent to the client yesterday that you may not be up to speed on. Here's your next upcoming meetings with these things and the most recent updates to documents. Like, here's a little one paragraph of what's going on with this client right now. If you'll take that as far as like drafting the first version of an email for me, trying to attach that tax return that I think they're asking for, all that stuff is gold from a strategic standpoint for practice management systems. It's interesting. When cloud practice management systems first came around, I think the main focus was task management. Like that was initially. The moat is if we can get accountants to manage all their tasks on our platform in a more modern way, then that's how we're going to keep them. But task management and cloud apps pretty quickly became table sticks. Like from your trellos to your clickups to, you know, now notion. And like everybody's done everything under the sun when it comes to task management and that's kind of table stakes. PM's next realized that it was actually the client experience that was the moat. Because if I couldn't get a firm to train a thousand clients on how to use the portal and how to use the client facing side of my practice management system, nobody's ever going to leave my platform because they're not going to want to retrain a thousand clients on a different portal. Right? So that client experiment, that client experience then became like the hot ticket. That was what everybody was building around and investing in. Now honestly, the next few years I see the moat being who's got the best email client. And a lot of, I mean 95% of accounts are still hustling outlook for their email, I'm fairly certain. But these like assistants that we're talking about are so helpful to never have to write an email from scratch again and for it to see into all of the other data in your firm and then draft or apply taking all that stuff into account and show you like, here's, I got this from here and this from here and this from here. And you can see that's why it's part of the, the draft email that it put together, that thing is going to be absolutely be your best, best friend. And if you fall in love with the experience of drafting emails within a platform, because you learn that assistant really well, you know, the things it does, maybe the things it doesn't do, but you've gotten used to interacting with your clients in that Email interface. I think that email interface might become the moat for practice management systems going forward. You are so in love with the assistant experience and how it helps you draft those emails that you can't imagine working another way. When we have these killer assistants that are helping us with email, you know how painful that is going to be to go back to the traditional way of writing email or to an email authoring experience that just feels more clunky. It'll be interesting, we'll see. But I honestly think like the core of these practice management systems going forward is client communications and who's doing that best. This episode is sponsored in part by Cloud Cloud Accountant staffing. Y' all know I'm a big advocate of hiring offshore. One of the biggest changes I made in my firm, we transitioned a legacy firm from 100 onshore local hiring to 100 distributed US and then 100 distributed globally hiring. And honestly is the best thing I we did it virtually alleviated all of our hiring pains, completely changed how we thought about staffing projects and the type of work that we wanted to bring on. Because you know what, the folks we hired offshore really freaking good. A lot of misconceptions around the type of people that you hire offshore because your enterprises will oftentimes use offshore folks for like menial work. Absolutely not the case. There are tens of thousands of people working for big four accounting firms. You know, offshore outside the US you can get folks that can do anything from tax to junior level stuff to super senior level stuff. But try to do that yourself, figure it all out yourself. That's going to be hard. It's going to be scary. Really good place to start, Cloud accountant staffing. They will hold your hand through that process. Their story is super simple. An accounting firm in the US hired a bunch of people in the Philippines, fell in love with them, but didn't fall in love with the fees they were having to pay to the staffing companies that were managing these employees. So they built their own solution and now they're starting to pull other accountants in. I'd encourage you. A big tipping point for me was when I was like, I'm going to stop being opinionated on this and just try to learn. And so I talked with other practitioners, I talked with some of the vendors that would like help you get into offshoring. That really opened things up for me. So if you've been on the fence, I'd encourage you to at least learn about it and start heading down that path. Consider Cloud accountant staffing next up. I think we'll see our first new entrant into the practice management system ecosystem that says they are an AI first company, whatever that means. There are like that there's some legitimacy to that being a thing. A lot of the things we're talking about the infrastructure required to like be able to intelligently suggest a document attachment, you know, an email reply and stuff like that. There's definitely some aspects of like how you build an infrastructure that's going to be best suited for that. And maybe that's a little harder for incumbent practice management systems to pivot to and otherwise maybe it's easier because they've got, they're able to invest in it in a way that other folks can't. But I think we'll see somebody a new entrance to this space saying we do this better than anybody else and that's why you should use us. Next up, useful advisory AI assistance. We touched on this a bit already. I think both in the tax space and in general in the accounting space we'll see consumer facing assistance that are pretty darn good, not to the level of a good advisor, but are a whole lot better than anything else right now. Because where the heck would you go to get advice specific to your accounting file or advice specific to your tax return? Right now there's way too big of a gap between a professional and Google and I think there some AI advisors will start filling in that gap and I frankly I think that's a good thing. Depending on the implementation like they could take it too far and lead people astray. But I am all for access. If AI assistance can enable more access to expertise, that's a huge win. Next prediction. I think the tax research software industry will be completely turned on its head. One of my I shared one of my biggest disappointments of this past year is we don't have people doing meaningful things with retrievals with all of the new technology we have when it comes to tax research. Supreme Court Justice I think who was it Chief Justice John Roberts just this last week said that basically all things legal are going to be completely turned on their head by AI. That court proceedings would be sped way up and be made much more cost effective by AI. That legal research will be completely changed and wildly simplified by AI. He said it's going to impact basically every aspect of the judicial system, which is great and he was very frank about things it does badly and things that you don't want to do. But when people say things like that like it raises the conversation to just a different level of legitimacy and all These things are going to apply to tax and the way that we do tax research and all that, you know, and, and who's to say what the future of, you know, court proceedings and all that is if we can get to a level of AI that can see into authority the authoritative information that it needs in order to actually make decisions. Like ultimately, do we want AI replacing judges? Probably not, but there's got to be a fast pass version where it will do a whole bunch of the work in a much more streamlined way that doesn't require tens of thousands of dollars in legal fees and all of that. Right? So once these research assistants are good enough to where they can see into all the authoritative information that they need, they're going to be able to come to like some pretty compelling conclusions on their own. And this is not the same as asking ChatGPT a technical research question. That's not what this is at all. This is using retrievals, this is pulling from the actual authoritative documentation and just using the text from those things to build out a string of reasoning and if you want give you an answer. But at the very least what it did was fetch all the information that's relevant to your question so that you can get to your own answer. And this is something that that AI can do and it will keep getting better at doing it. Right now it requires put pulling together all the, all the underlying information that has to go into being able to answer a question like that. Like this is a complex step by step thing that I don't want to, you know, make it sound like that's really easy to build. But I'm a little worried at how slowly the tax research companies are moving. These are not super progressive companies generally. I've said recently, right now Google Bard is probably the best tax research platform there is because it will do like its own built in fact checking of all the stuff that it returns. The biggest downside is you can't give it access to a huge library of authoritative stuff. And one way or another this is going to happen and will will completely change the price of that access as well. Right now tax research products are wildly expensive. I think AI will bring that price significantly down and more importantly bring down the cost of getting to answers by a ton because you can get to those answers much quicker. Quicker using AI to, to make you more efficient. Next prediction at OCR is solved and we're like, we're pretty much already there. It just hasn't been implemented into, you know, hardly anything. So right now, you know, OCR does A reasonably good job of reading computer written things, but there's so many things that fall outside of that and so many bad scans that it's not going to read well. And all that compared to like GPT4's vision model where, where it will take an image and can read just about any sort of handwritten stuff at the level of a human being, if not better. And like that is legit. That is the truth. The problem right now is using the vision model to extract every bit of text in existence is much more resource intensive than ocr. So if you're building this into a software product right now you gotta figure out well when do I use ocr? Cause that's good enough versus when do I go full blown like use the vision model and, and rack up my OpenAI bill. So the exciting thing is we can lift text off of a piece of paper as good as a human right now just using machines. What we're waiting on is what does that look like being implemented into an attacks workflow? I mean if you think about, you think about take and I won't go too far down this path because a lot of non us and non tax folks listen to this. But take a 10, 40 and 50 pages of unstructured information. How much organization goes into putting that into a useful document. You've got, you know, services like Gruntworks and PFX Scan and all these, all these different things that will put it into a bookmarked PDF for you. Ta da. Great. Wow, thanks. But then like you've got another app to like extract information out of it and then like put it into your tax software and if you want they'll have an offshore team or an onshore team like review the output for you as opposed to an AI enabled tool that can literally lift all of the text off of that through a combination of OCR and vision models and then organize that into like a structured work paper, like read the handwritten notes from the clients, like pull the emails from the clients from your email system. There's just like a. Once these tools are rebuilt with like the best AI tools and tech that we have right now, they're just going to be fundamentally better. Three more predictions. First, Microsoft Copilot will underwhelm and continue to be behind the best of breed tools. It's pretty big bummer. All things like Microsoft Copilot right now are just not frontier, not the best of the best. And like every single application there are other tools that are better than it. So the reality with where Microsoft is right now is they're just behind. And the biggest upside of Microsoft is them being able to roll out a killer thing that can see into all the data of yours that they already have. So does a tool have to be the very best of the best for that to be really useful? Probably not. Just being able to see into my calendar, into my meeting transcripts, immediately makes the assistant more useful. The frustrating thing is when you know there's a better tool out there for the job, it just can't see into your data. That's where we're at right now. And it's really frustrating the fact that you have to go out to ChatGPT to, you know, have it write that Excel macro for you or something like that. And that's going to be way better than, you know, Bing Copilot or whatever Microsoft is calling it. So I don't. Whereas, like, last March, we were all pinning our hopes on Microsoft building, like, this assistant of our dreams, I've definitely cooled off on that. I mean, reality has been almost 12 months now, and we, we, most of us still don't have that in our hands. If they had an unbelievably good version of it, we would all be using it already. Next prediction, I think we're going to see more effort being put into consolidating our chat assistant experiences. So, like, Windows Copilot is built into every Windows desktop now. There's an ecosystem where you can build plugins for Windows Copilot. And what none of us want is to have to use a different chat assistant in every single app that we use. Right? Like, that's just going to be terribly annoying. Eventually, I think we see the consolidation of a lot of this stuff down to a more unified chat experience. So using the example of tax research, if I can get a ChatGPT plugin that will do tax research, good enough for me. I'm just gonna use that, man. And I'm never gonna use ChatGPT, right? So, like, there's entire categories of app that I think will be able to ultimately be plugins and thankfully enable us from having to go out to a different place to do that work. And now we can do it just in the chat experience that we prefer. And then it'll work with stuff like, you know, speech, like being able to talk to ChatGPT. So if I want to work on that thing when I'm on a walk or something like that, it can lean into those other functionalities. If it's also connected to my QuickBooks file or my email, like, then I can kind of bridge those different data sources from that single chat app. So I'm hopeful we'll see more effort into like consolidating our chat experiences so that we're not having chat experiences in every single different app that we use. Last, as a biggie, something significant will be eliminated. The biggest thing yet. Is that vague enough for you? One thing that it could be that we already talked about reconciling bank statements and QuickBooks Online. That's a great example of something where when we hit an itty bitty tipping point where it's one day not good enough and then the next day good enough, it's going to be a big freaking deal. And let's, let's say it's that because we're actually pretty close to that right now. Being able to reconcile to a bank statement to like tick off that monthly reconciliation inside of QuickBooks Online. Somebody's going to figure that out. It's probably going to be in a Chrome browser. And when we get there is going to be a really big deal because there's certain really high volume things that we all do where as soon as we get an AI agent good enough to do it, it's going to be like oh my gosh, I'm going to be able to use this 100 times a month and it's going to save us time. I think frankly we could get that with like 1040 prep, like managing simple 1040s. At least getting it to a spot where you've got like a work paper that's tied out and can proof to like a mock up of a 1040s so that it like says okay, this is what, this is what your tax software should show. What's that killer Excel workbook that guy makes every year? That's it's like a giant work paper for 1040s and it even has like some checksums for like what the tax forms ought to ought to show. I could see AI tools that can like knock out the, you know, low hanging fruit like simple tax projects like that. But I think we're going to get something that's like a meaningful task where it's the first example of an AI agent like wiping out this thing that is one of those like core tasks for accountants. And for me that's exciting. What I always come back to is like when I was running an accounting firm and I had a team of people, if somebody put this new tech in my lap that just wiped out this really time consuming thing for us, I would be so jacked because Every single person on my team, I had better things for them to do. Like, every single person I'd have a conversation with, it's like, I get you're doing this right now. I really want to get you to this. And if overnight that thing got wiped out, that would be awesome. I'd be so pumped. Now if we weren't in a position where, like, hiring was scarce and we're always trying to find more accountants, maybe that wouldn't have been the case. Maybe we would have let people go. But, like, that was just in our, in my situation, there was always so much more to do than we had time to help with. And so I was always so happy for any of my accountants to get time back. So I have a hard time seeing job displacement being much of a thing in our space. I think it definitely will be in other spaces, but this year I think we'll get our first, like real concrete example of a thing just being like, poof, that just got 90% better overnight. Maybe that's tax research. Oh, I'd love that. That's it for the roundup. What do you think? What, what do you think will happen? But also, is there anything that you would love to happen? Last year, Nio put out a tweet now Carter Gray, that was like, if you could have an AI assistant of your dream knock out anything right now, what would it be? And obviously it's accountants. So everybody's got their snowflake super nuanced, trying to out think everyone else answers. But like 50 to 60% of people said email. And so that's why I always come back to email as the biggest opportunity for tech vendors in our space. Build a product that solves email with AI in a really compelling way and people are going to love you for it. That's probably still the thing I'm most excited for, is just like AI to tame the email monster. But got anything you super want to see? Drop them in the comments. Disagree with anything here, let me know. And here's to a magical 2020.