Transcription
The new GBT5 AI model, it's not just smarter. Accounting firms are already putting it to work, using it to create their workpapers, to review their workpapers, to do actual bookkeeping. Not my QuickBooks. They're even using it for like sensitive tax stuff, which, by the way, can we even trust AI with sensitive client information right now?
I last ran a 40-person accounting firm. I can tell you today, the stuff we're running through would have changed how every member of my team worked. Today, I'm going to show you a QuickBooks use case that you're going to love. Some research ideas that will save you a pilot time to Excel use cases that no previous AI model could do. And my very favorite one, using ChatGPT as a full-blown, like, I don't want to spoil that yet. Let's just say things are changing. You're going to love the last one. Okay, real subtle YouTube, boy.
Now, importantly, I'm going to share some AI guardrails for ChatGPT5 because if you misuse it, as many accountants are right now, it's something that can actually put your entire accounting firm at risk. But if you get it right, buddy, you are going to be pumping out work like nobody's business. Okay, let's get into it.
QuickBooks use case. Client's got a brand new loan, $25,000 for a truck. Yes, room, room. Here it is. Notes payable, $25,000. Client sees a brand new ride. I see five years of adjustments, allocating principal to interest. But what if ChatGPT could post all the journals for every payment right now? Let me show you how. It's pretty simple. And don't sleep on this use case because it's actually a gateway drug for the next one that'll handle your most wildly complicated payroll journal allocations.
But first, look at this. It's an amortization schedule printed off a super legit calculator. We're going to chuck this bad boy into ChatGPT. You're going to help me create some journal entries I can import into QuickBooks. Attached is the Excel example journal entry they give you. Also attached is my company's chart of accounts. And here's a help article about best practices. The QuickBooks article about how to import journal entries. This is coming off. Okay, there's so much good stuff in this video game.
[Music] Close.
Now, you notice up here it says ChatGPT5. And with this new model, they have simplified the model selector. You used to have to have like a PhD in living with your mom to know exactly what model to use. Now, by default, it'll go on auto. There still are ways to go back and like unlock other old models. I'll be honest, I still don't trust the auto picker. I usually just use "thinking" if I'm doing accounting stuff. If I'm doing creative stuff, I will switch to "instant," but I usually just leave it on "thinking."
Now, let's finish out this prompt. The attached loan originated on September 1st. I've set the payments in QuickBooks to go to a loan account, but generate the journal entries to move the interest portion over to interest expense on the last day of the month. Let's do this for all the entries through December 2026. I'm going to kick that bad boy off. And you'll notice it is thinking. That's because we're using the thinking model. Take some more time to make sure that what it's doing is correct before it comes back to you with a response.
Now, over on the QuickBooks side, we can come up here to settings, say import data, journal entries, import. And here is like a sample CSV that they give you for how to like structure the files that you want to import.
Okay, it took about two minutes. I can download the CSV here. This is the thing that we can import into QuickBooks, which means anything you can get ChatGPT to create for you can be pushed into QuickBooks with a journal entry. Same story for Xero and just about any other ledger.
Now, take a quick peek at this in Excel. Is that Excel for Mac? I knew this guy wasn't a real accountant. You can see this is pretty straightforward. You've got a bunch of journal entries here through the end of '26, which is what I asked it for. Pop back over to QuickBooks. Grab the CSV that ChatGPT just made. It figures out all the column headings on its own. You don't need to fuss with that. Complete the importing. It tries to link you to the journal entries, but it usually breaks. That's the QuickBooks Online that we know and love. Let me go to the journal report. Make sure that posted. Yep, here's all the journal entries from November 30th. Let's see. So, let's double-check our amortization amount. Yeah. So, $83.33 on the first month. There's $83.33. Let's just jump down to December, $64.03. Yeah. $64.03 right here. The 16th payment. Just like that. How you like them apples?
Okay, that was like the foundation for the even cooler stuff we're getting into now. How to use the same thing to import even your hairiest payroll allocations. Get AI to build the whole journal entry. Then we're talking research. Then we're talking the the, honestly, my magnum opus. The use case I think is so cool. By the way, are you on my weekly newsletter? Over 10,000 accountants are on that. We're sharing rad AI use cases like this in there every single week. It is like five minutes of reading every week that will ensure this AI stuff doesn't pass you by.
Okay, payroll journals. This one's going to look very similar. Exact same prompt. You're going to help me make some journal entries I can import into QuickBooks, but I'm just going to add, "Create a journal entry import for the attached 92 pay run." And it's this payroll journal report from my payroll system. It is a real beauty. These are always super easy to read, right? I'm just going to upload that to this conversation. That payroll report straight out of the payroll system and I'm going to cut it loose. Remember, it's got my chart of accounts so it can infer what that entry ought to be based on that payroll journal report. But if there's some nuance there, you can actually give it like the last journal entry and it'll follow that one to a T. But in my testing, this actually does it totally fine on its own.
Now, it took this one about two minutes to build me this CSV. But it didn't use all the accounts that I wanted it to. It just used the parent payroll expense account. And so what I did is I actually gave it a screenshot. I said, "Use the detailed payroll expense accounts." You see, we've got like 401k contributions, dental insurance, salaries and wages. I said, "Use all that, not the parent stuff." It worked on it for 45 more seconds. It updated the CSV, and you can see it hit all the accounts I wanted. It's doing this intelligently, too. It's only pulling the employer portion of these things. It's dumping everything into payroll liabilities, both the employee withholdings and the employer portion. The only payroll liability account I have in the chart of accounts is payroll liabilities. But I've done like three to five permutations of this in testing. It has nailed it on import every time. That was that was like five employees.
"We're not impressed yet." Here is where we jump from 200-level to 400-level, my friends. With ChatGPT's very best feature, agent mode.
"Check this out. Becky, we're doing the thing. Becky's got a spreadsheet she is very proud of. Almost too proud. Here it is. And let me just zoom out here. Okay, now you're speaking my language. Basically, what this does is it takes all of the people and allocates them across a bunch of different departments. You know, we have there's we have so many dumb versions of this in the wild, right? Be it the Amex statement or the folks on payroll getting allocated 10 different ways, then the owner wants to fiddle with it. Basically, in short, and I know this is very small, you put the employees here, you put the current pay run info here, and it gets allocated across departments. It's not rocket surgery. What if we took this little benchmark from five employees to 30 employees? But this one's a little different. Here's a work paper template we used to calculate a special payroll allocation and we attach it, update it with the data from this pay run, and we'll now give it a payroll journal for the September 2nd pay run for those 30 employees. Now, make sure you got agent mode enabled. We're going to kick that off. And agent mode, they actually said this in the announcement. It is the most powerful AI we have ever had for doing spreadsheet stuff. It's particularly good at taking spreadsheets that you give it and updating it, putting data into it, and it honestly remains my single favorite feature in ChatGPT right now using agent mode to do spreadsheet stuff. This took all of four minutes. Here's the updated payroll allocation work paper with all the goodies. That, that's it. You download this. Check it out. Can we get some kind of drum roll? I'll drum roll that. Okay, here we are. It put in all the data for all the employees. And this is a great use case because it's easy to double-check your numbers, right? So, let's open up that payroll journal. Uh, come down here to the bottom of gross pay. Actually, let's just sum it. $84,320. Total gross pay here. Enhance $84,320. It's doing some like really intelligent stuff here because that payroll allocation spreadsheet doesn't map perfectly to the payroll journal. So, for example, all the employer-paid stuff, it's this column, and that's the sum of employer FICA and Medicare, unemployment, employer share of 401k contributions, insurance, it's all of this stuff. It's $18,070, the same as this total. We flip back to this, and the input fields here are employer payroll tax and employer-paid benefits. Scroll down here to the total. It's that same $18,070. It is figuring all that out for all of these entries. How many entries? The count? 150 entries it made here from probably over 200 source figures on the payroll journal report, gang. It did that in four minutes.
I'm coming in now. Let's take a big brain break here. Let me give you my big brain tip number one. We just saw there were several prompts that I reused several times there. And it's important you save these prompts that you build, potentially even share them with your team. A really easy way to do this is with text expanders. One of the most popular ones is called TextExpander. Basically, you type a little command like semicolon prompt, and then when you do, it drops the whole text in there. We've shown this in some past ChatGPT examples. But if you run a team, the beauty of this is you can build out your favorite prompts, but then share those prompts with your team who also use that TextExpander. Then when you make updates and improvements to those prompts, the snippets are also updated for your team members.
"What, what was that one called?"
"That's TextExpander."
"TaxExpander."
Okay, some practical advice of using GPT5 for research. It can be a massive timesaver here. And this isn't just for tax folks because even bookkeeping clients will regularly ask us, "I just found out I have to complete a quarterly B-census. May you help me with this? Do you even know what it is I do?" Thanks so much. I'll do a tax research example here, but same principles apply regardless of what you're researching.
Prompting with GPT5 looks a little bit different. As I mentioned, it'll try to figure out what the best model is for you. But myself, I'm not going to let it do that. I'm going to force it to use the thinking model. And if you noticed there, it took off on its own very, very quickly. That is because it didn't use the thinking model. So, let me try again. A new chat. I now have "thinking" selected. I'm going to put the same prompt in there, and you'll see the difference. So, it gives you a little pause. It actually says "thinking." It is a little bit slower, but in the grand scheme of things, we aren't talking about all that much time.
Something GPT5 is much better than past models at is knowing when to go out and fetch like external information to support what it's telling you. The biggest thing to look out for when you're using AI for research is making sure it didn't just give you a response from what the model knows. Because these models are trained on like vast amounts of information, but it's information from every tax year, all sorts of different tax jurisdictions. So, I don't want it to give me the average of all of that. I want it to actually go out to a source, gather a bunch of information, and then give me an answer based on that source. The trouble is the sources, as you'll see here, aren't always the sources you wanted to use. That's not true. This did the research perfectly. It's only pulling from IRS sources. Okay, goodbye.
So, look at this prompt here. I called out 2024, and that's very important. Rules change year over year, right? But then I also called out like the rules body. And if we want to nitpick, like, no, these are not authoritative sources. They're publications. I can tell you the better version of this prompt, which was how this demo was supposed to go. It's the same prompt, but you say, "Only pull info from IRS.gov." The last one wasn't supposed to work. Okay, but it's really as simple as that. Otherwise, it's pulling from like NerdWallet. In fact, that non-thinking answer that it gave us, let's go back to that. It's a great example of the downside of not using thinking potentially. So, look at the sources that it pulled for this. We got NerdWallet. We got the IRS. That's all right. Small Business Trends, Investopedia. I know about you, that's where I always start my tax research. Business Insider. Okay, see, use the thinking model just by toggling this little drop-down and go to "thinking." And don't be afraid to tell it, "Only use sources from IRS.gov."
But what is now possible with GPT5? Because honestly, past models did research pretty well. You can now ask it to do more complex work with this information. So, I'm going to give it a meeting transcript. "Meeting with Clyde. Analyze the details from this transcript to determine if my client could claim a home office, then break down their different potential offices and the criteria for whether they would or would not qualify."
So, home office, you can actually have several home offices that are eligible. I think you combine them when you report them. I don't know. But you'd have several places on your property, and this is an example of like a multi-step request where we're saying, "Analyze the details from this transcript, then go through each potential office they mentioned and break down the potential eligibility of that office." And this is a level of effort that previous models were not capable of.
"Did he just put a sensitive client transcript into AI?" Sure looking that way.
Okay, it spent two and a half minutes. Bottom line, he can likely claim a home office deduction for the locked Bureau of Information Annex. The taped-off half-bedroom doesn't qualify due to non-exclusive use. And the second bedroom may qualify under the inventory storage exception if it's regularly used to store products. This is important. We're actually going to get to this, how usage could be modified a bit to make it eligible. It breaks down kind of the requirements space by space of what it would take for it to be eligible. Again, gang, this is so valuable. This is such an incredible timesaver.
But here's the thing. Can we get to big brain tip number two? Is this is only as useful to you as it is provable. This is an intern bringing you some work. You ask them to go research this and they came back to you with something. When they turn up with that, are you any closer to an answer than you were before? In some ways, yes, in some ways, no. It's why we teach our team members to make the work that they do easily provable. So, we're going to come back to ChatGPT here and say, "Dig into office number two further. Give me proof of your findings using exact quotations from the IRS is important to support each point and calling out any related exception." If you have ever done tax research, you know, you think you got to the answer and then seven pages later you're like, "Oh, here's a list of exceptions. Cool." It's what makes tax research so time-consuming. But gang, it is also what can make technology such a timesaver because there's a lot of people right now that scoff at this. They're like, "You can't use it for that." And there was a day where it wasn't good enough to be useful. That day has come and gone. It is such a fast way to get to the most relevant information.
Okay, here's the response. All right, a deeper dive here. Why it fails the exclusive use test. Exact quotations here and then the bullet points of the facts from the transcript. This was actually the thing that jumped out to me, the fact that you could make a half room eligible. Yeah, this bit it doesn't need to be marked off by a permanent partition here. The only exceptions and whether or not they apply. I would say right now you still got to approach this stuff like this is a junior bringing you information, right? But I would say that the quality of what this brings back is really good for something that can work, you know, in minutes. How long did that take? Three minutes. And I would encourage you as you're getting comfortable with ChatGPT for research, think, what is the easiest, most useful format to get this like information back for me? How would I ask an intern to do it and then have ChatGPT do it the same way.
Now the really cool thing here is we can come back and say, "Follow up with suggestions to make these offices eligible." Because sometimes it's as simple as moving tubs from this room to that. And when I did this earlier in testing, it went as far as to say like, "Hey, send me a link to their Airbnb listing because it's got like a couple locked rooms that he could use for storage," and it would suggest changes to language. So, you can even like give it a link to the person's actual Airbnb listing. This is completely irresponsible. What am I suck?
Okay, we made it far enough. Let's talk security scaries because there are still completely irresponsible ways to use AI. But many of the fears are founded in information that is now out of date. Let me take you to the ChatGPT business landing page. Any AI tool that you use, this is what you need to look out for is this right here. Are your prompts used to train the model? Because back when ChatGPT first came out, every single thing that you put into it was trained into it. And that is still the case on the free ChatGPT plan.
"Yeah, honey, I need you to hop onto my computer real quick." But the fact that there remain insecure ways to use AI does not mean that there are not completely secure ways to use it. Microsoft, Google, ChatGPT, they all have business plans now that do not train on your data. Which means security due diligence now looks the same as it does for any other AI app. But you still see a lot of this problem is no one knows where the data is going. Can we do it with client's data without their consent? What about compliance? The biggest issues with any online AI is it's not private. You're literally sharing your details with anyone, feeding the robot data to quickly replace you. Smart. Wait, you're sending all your client data to Sam Altman? The reality is the way AI works today, it is well understood. We're still figuring out the cool things it can do. But the way it uses our data, there are no questions around that. It's why please, please, please, I recommend to firms get everybody on your firm onto ChatGPT for Business. Comes with the data protections you want. It is AICPA SOC 2 Type 2 certified, which is a higher bar than many commonplace apps in the accounting space. Enable multifactor authentication and do this now because if you don't, I can almost guarantee you that there are people on your team that are running out and using the free version, especially if they are, uh, 23 years or younger because we got a whole generation of folks that just came out of school using ChatGPT to do everything. And when they enter the workforce, they're going to be freaking wizards. They'll be capable of doing things that like you and I cannot do. But we got to give them a responsible way to use AI. And I get there's still like this newness around it and you're like, "Ah, can I actually trust this guy?" Man, go work with whoever your IT help is. Try to get really first principles about this stuff because this is being used in all sorts of unbelievably sensitive applications and it's a huge opportunity for us and our firms right now.
Now, I want to show you a couple Excel examples that no previous AI models could handle. But first, a word from this video's sponsor, Rolling Music. Buddy, it's Carbon, the practice management system. In fact, they've been one of my top recommended platforms for years, and they were the first in the game shipping really cool AI stuff in their platform. Look at what Carbon AI can do. Look at this little conversation summaries. Like, bless this feature. There's a little digest of everything that's going on with a client. So, if you're tuned out, you can hop back in and see like a super quick AI-generated summary of what's happening. A big part of Carbon is email triage. They're going to use AI to help you assign emails to the right people in your team. These two right here though, these are actually my favorite. They have built-in AI in the like email client to help you do like a quick reply to a client's email. It can see like all the email history, like the previous replies, or even use AI to write something new from scratch, which that's me. The release of Carbon AI this year set the bar for how practice management systems should be implementing AI. It really does. And I will tell you the folks at Carbon understand that our AI will only be as good as the context it has access to. So the way they're kind of building their platform right now is they will be the home of as much context as possible for your accounting firm so that the AI then built into the platform can do really interesting and novel things. And they're delivering. Like they just shipped a first-of-its-kind deep integration with a tax workflow tool. They now have an integration with Vynyl, a meeting recorder tool. Imagine all of your meeting recordings coming back into your practice management system automatically. But then like when you got all that context in one place, you could see a future where AI is able to unlock some really cool stuff for how we manage our firms. Now, Carbon's got a very cool State of AI report they put out every year. Practical stories from firms on like how they're using it right now. Some really interesting data. I'd recommend you go down and grab that. We'll link it down in the video description. And thank you to Carbon for sponsoring this video.
Now, a couple Excel use cases that could actually save you a ton of time this week before I show you just the most mind-blowing one that I love. Now, look at this. I call this the payments into deposits benchmark. Ah, wow. Okay. Payments here, deposits here. Gang, we have all lived through some version of this. You got a pile of numbers and they roll up into the other numbers. We did a bunch of stuff with dentists. Oh my gosh, was this a mess to unwind because these numbers, there's so many different combinations of the ways these numbers can go to add numbers. And it's something I've always used AI to try to solve in the past. Couldn't do it until now. Check this out. Two sets of numbers. Here's the payments. Here's the deposits. It thought for five minutes because there's actually an insanely like exponentially large number of potential combinations here. It runs through each of the deposits and what makes them up exactly, but then it calls out three that it can't match perfectly. And that, that is the correct answer. So what actually makes this so difficult is there's like it's not a foregone conclusion. It will match just right. So here's the solution. And these three don't have perfect matches. And it got that in five minutes and actually proposed an alternative that I didn't even know about. And I like this came up for me in like accounts payable applications in some e-commerce applications. And to be able to just like copy-paste it in or grab a couple screenshots and be like, "Make these go into these." In fact, very similar. I call this one the "I can't be bothered to figure this out" benchmark. We got a list of transactions in here and a list of stuff on screen here. But the totals are off. There's some missing number or a number's off and there are so many in the list I cannot be bothered to go through it. Look at this. Your new best friend, ChatGPT. I tried to key everything in but I must have missed something because the totals don't match. What transactions here have different amounts or are missing? So I gave it this CSV and just the image, the screenshot of the app. Worked on it for about three minutes and it said three transactions are different. This one's a different amount and these two are just straight up missing. Like that, gang. Three minutes. Not the whole tracking down the, oh, do the, the ruler, uh, get to the end of 100 numbers and cross your fingers and hope you got it this time. No more. Literally chuck a screenshot in here and it can do this. I had a hundred use cases like this that like GPT4 past models have not been able to do.
And if I can give you my big brain tip number three, buckle up for this one. My best hack for you to use ChatGPT more is to leave it up on one of your extra monitors. The biggest blocker to us using it more is remembering, "Oh yeah, that's a thing. This could help me. This could save me time." And so if you put ChatGPT up on one of your displays, like close Outlook, man, like put it up over something that would otherwise distract you and you'll remember that it is there, just just waiting eagerly, wanting to help you.
Okay, this has all led to this. I teased my magnum opus. Is it magnum opus? I really hope it is. I'm going to get made fun of. It is magnum opus. Okay, good. What if ChatGPT? What if it's specifically GPT5? What if it was capable of functioning as a full-blown accounting ledger? Okay, hear me out. So, I I'm convinced that every accounting firm has their own version of this spreadsheet. It is like an accounting ledger in a spreadsheet. Basically, you put the transactions in here. You got a starting balance up here, $100. $50 transaction, and then in here, you put a letter and it codes it to that column, right? So next transaction here, code that as software, puts it over here, and and you fill in all the transactions. By the end, you got like a very basic P&L, right? We all have some version of this. In fact, the first like five years of my accounting career, I was doing trust accounting in a spreadsheet like this, entering stuff off of like bank statements. My golly, by the way, got bank statements? ChatGPT, it can extract those transactions flawlessly for you. But remember before when we showed agent mode? And we saw how good it was at doing spreadsheet stuff. Ledger template AI that can very capably do things in spreadsheets. Look at this. This is so stupid that this works. Nobody would have told you the future would look like this.
I use the attached ledger template to do basic accounting for my clients. Basically, you enter all the transactions, then just enter a letter to code the transaction. So, you only modify columns A, B, C, and E, and the rest are locked. Now, put in all of the provided transactions and perform an initial pass of classifying via column E. But come back to me for clarification. Like a list of those you're unable to categorize. I mean, this won't work, right? They can't do this. Large volumes of data entry. Now, enable agent mode. Cut her loose. $20 says it works. Is this going to be a problem for us? Only two minutes later, it comes back. "I've loaded all 50 transactions in, but I need your help with these guys." I fill in the blanks. Spotter is software. Charger's office expense. PayPal's office expense, as it always is. 14 seconds later. Here you go. Here's the final version. Set your peepers on this one, gang. All 50 transactions categorized. And it does a pretty good job. Like no worse than my numpty intern. This is easy to foot also because we can check the running balance against the statements. But that's just 50 transactions, right? What if we did 419 transactions? Here's the big boy. Here's the actual file I'm giving it. What do you even format Excel? Anyways, over 400 transactions. Are we giving it all 419 transactions? Enable agent mode. Uh, talk amongst yourselves. It takes a, it's done. So, it took four minutes and it came back with this big old list. It said, "I don't know what these things are." I'm like, "Buddy, you know that Uber is travel." So, I said, "Bad robot. You can figure out more of these." Like, Uber is obviously travel. Work out some more and tell me what's left. One minute later, let me show you the list it came back with. Like a junior that has been properly scolded, it has now taken some creative liberties. Now, I told it what each of these transactions are. 30 seconds later, bada boom, bada bing. Here we go. Open this up. 400 transactions. Every single one with a classification. Just now realizing this formula is broken. But don't worry about that. 400. Think about that, gang. It entered the date, the merchant, the amount, and the class for over 400 transactions. That's over a thousand entries that it made in less than five minutes. I can put the total like that because I've got a running balance. Less than five minutes. And frankly, it has me rethinking the best ways to use ChatGPT because any process we can now build in a spreadsheet that's got the built-in checks and all that that we can then hand over to ChatGPT and say, "Do this." I'm talking repeatable processes, stuff that we do a lot that just changed the game, right? Too hyperbolic? I don't know, man. But agent mode, it is a big, big deal. But you got to watch this next video to learn more about it. Importantly, some context where you should never use agent mode. Like, did you know it can work all by itself online? Like, you can cut it loose in a QuickBooks file. Is it too powerful? Check out this video where we talk about the boundaries of where to use it and where to steer clear.
[Music]