498 First Impressions: GPT-5 For Accounting Firms

Stop what you're doing. There's a new really good AI model in your chatGPT. OpenAI. They announced and actually launched GPT5. It's a big deal. Some folks are very excited, A lot of people are very mad. We'll get into that. But it marks the first time we're at like a entirely new number increment of a GPT in over a year. But does it deliver? We're running through three like real life accounting firm use cases to see how it stacks up to the previous models. How big of a step is this? But then I've got three other thoughts for like how this is kind of reshaping what I'm thinking thinking in. In the near term how this impacts accounting firms specifically. So let's get into it. Everything accounting firms need to know about GPT5. If you're new here, welcome to the Jason on Firms podcast. This is a podcast, so you won't be seeing me demo GPT5. You want to see more visual demos, check out the Jason on Firms, the main YouTube channel. If you're watching this on YouTube, we got an audio feed as well. But this GPT5 for your ears since most people listen to this on audio. So for context, GPT5 it is a new generation of AI model from OpenAI. The folks behind ChatGPT. OpenAI historically awful at naming their AI models. Back in November of 2022 and ChatGPT first came out the model behind the scenes there was GPT 3.5. The following March we got GPT 4 and just now we are getting GPT 5. How long is that? That's. That was spring of 23 we got GPT 4, so over two years later. Took over two years to get 5. Nobody would have guessed that. But in the middle there we got other models like GPT4. O O I think stood for Omni and it was like an upgraded version of four. Even though these models like even the ones with the same names, every couple few months they actually get upgrades. So there's like a March version of the model and a May version of the model. But just like the confusion around all the naming, completely understandable. They went from 4 to 4. 0, then launched 01. And that was the first unquote reasoning model. That is the model will actually think about its answer before it comes back to you. And that was beneficial for a number of reasons. Kind of like talking to a person and you ask them a question and they give you an answer that's just a little bit too fast and you're like there's. How did. I don't trust that you actually knew that off the top of your head. These thinking models, they would do just what the name implies. So we had 01, then we had 02 and then we had 03. So before last week inside of your chatGPT you very confusingly had a model called 4O which was older and weaker than the model called O3. And I can't tell you how many people, in fact I'm coming off a run of several weeks of accounting AI like live summits. I can't tell you how many people were like man, AI sucks at X, Y, Z. And I'm like eh, it ought to be able to do that. And I asked them what model they're using and they're like the newest one four zero. And I'm like no, that one's. That was actually a lot worse for technical accounting stuff than O3, which is a thinking model and newer. And so OpenAI's solution to all of this is supposed to be GPT5, one model to rule them all, which actually isn't one model. It is a whole bunch of different models, different models with a built in, what they call switcher. So the idea is you can rock up, talk to GPT5 and it will sort of assign out whatever you tell it to do to like the model that makes the most sense. GPT5 or GPT5 thinking, which is like the reasoning version of it. But they knew the naming of all this stuff was super confusing. And that was part of what like that was a big problem that GPT5 was supposed to solve is the simplification of not having to know what model to use for X, Y and z, thus making ChatGPT a little more accessible for non experts. But in the process of doing this they took away access to all of those previous models. So you're 4.0, you're O3, you're O3 Pro, you're at 4.1, 4.1 Mini. All these old models, poof, overnight they disappear. You can't use them anymore. And so all the hardcore nerds are ticked right now. They're like you just did a rug pull of eight old models and instead all I have is this new model which right now has some issues. The whole like switcher on day one actually wasn't working at all. And right now it can kind of be frustrating. Frustrating because it feels like you've lost some control. It's deciding things for you that maybe you wanted the ability to decide for yourself a great example. One of the example use cases we're going to go through later, I ran it once and it thought about it for two seconds, gave me a bad answer. I immediately ran it again and I thought about it for four and a half minutes. Gave me a great answer. Another example, Sam Altman tweeted an example of him like prompting it to make like a song about the GPT5 release with it's. I think it's called Beatbot. It's like this drum sequencer that's evidently built into ChatGPT and I'm like, I didn't know that was a thing. So I pulled up chat GPT on, on iOS. I put the exact same prompt in that he showed and GPT5 responds, I don't have access to, to Beatbot, but I can tell you how to go out and use it. And I'm like, what do you mean? Sam Altman literally right here says he can do it. I then tried the exact same prompt in the browser like chatgpt.com and it worked. So frustrating. Right now seems like you get different results depending on where you go. And this is a temporary problem, but this whole thing like we've been building up to GPT5 this new amazing model and we want to have this big sort of ta da moment when it drops. Right now it feels like that's really been lost due to like the switcher inside of GPT5 not doing what we want. They've already said they're going to bring back some of the legacy models and give you a little bit more control over picking what model it uses. So they're already kind of caving to the more hardcore folks that want a greater degree of control. Hopefully in the coming weeks we see some rapid improvements here because it does impact the overall usability of it and kind of be a frustrating experience at the moment. As I'm recording this, you do have the option to pick between GPT5 and GPT5 thinking. And for all but quick, like sort of creative brainstorming for accounting firm stuff, you will almost always want to use GPT5 thinking. That is where it'll take anywhere from, you know, one to 10 minutes to kind of process the problem and give you a result. And most examples I'm going to give you here they are like actual accounting, like work. And for sure we're going to use AI for other things, like all the things we are not expert in, like give me a marketing plan for X, Y, Z or help me rewrite this. And for many of those things, the non thinking version of GPT5 will be better, but for like the real technical stuff where we want to make sure it's giving us a right answer with no hallucinations, you will almost always want to use GPT5 thinking and you can actually prompt it like some of the examples were. If you tell it to think long and hard about this, it will almost always use thinking but you can also at the in the model selector just force it to use thinking and I don't know, I just don't trust it to like pick that correctly for itself yet for whatever reason. But the highlights on GPT in the kind of benchmarks, the big brain benchmarks of how good can it be? It beats all the previous stuff as you would suspect, but interestingly it's only marginally better than ChatGPT's agent mode. So we talked a few weeks back about a big release, this new agent mode that has like a built in browser. Previously we had a version of something similar called Operator, but it wasn't really all that great and it was only available on the 200amonth pro plan. Now on all paid plans, notably the team plan, which if you are an accounting firm using ChatGPT, you want everybody on the squad using the team plan so you can create a team plan account and then add members of your firm to it comes with all the data security required for using this with like actual client information prevents things like being able to share conversations externally. If you saw the headlines the last couple weeks about being able to find people's ChatGPT conversations on Google and how it's like this big ChatGPT security breach is not the case for anyone using a ChatGPT team plan. Please don't cheap out on this. Get your team actual get the folks in your firm ChatGPT team licenses so everybody with a team license you've got access to this agent mode which on the benchmarks is still almost as good as GPT5 and potentially for accounting workflows might still be better because one thing it's particularly adept at is spreadsheet stuff. It was actually a big step forward than anything in the past at like working with Excel files. And I've actually got one example we're going to go through of where I think agent mode is still superior than GPT5 and all that means is when you're inside of Chat GPT, if you're giving it a prompt, you just click the Tools button and select Agent mode. There's a little toggle there to enable Agent mode, yes, there's a little plus button and then you just enable agent Mode. But otherwise GPT5 is the best on all the, all the benchmarks. But interestingly, the way that they introduce it on the GPT5 landing page is they say, quote, it is the strongest coding model to date and writing code, it is still like the biggest, most powerful use case for large language models, arguably. And like, with each new model, what they're actually benchmarking is its ability on like math problems and writing code and science and stuff like that, because those are more deterministic sort of things where you can have a question set and a very clear answer of is it more right or is it more wrong than the last model was? But for accounting firm stuff like there, there isn't a clear benchmark of what's it better at, what's it not better at? And so a model that's better at coding it doesn't necessarily mean that it's any better at accounting stuff. And even accounting stuff is so broad because there's so many different jobs that accounting firms are tasked with. I didn't actually know if this was going to be any better than past models. Initial impression is I actually do think it's a pretty good step we're going to get into that. So candid accounting, free test tests, stuff that I've done on past models that I want to run through this new model to see how it works. So test number one, just give it a balance sheet, a profit and loss, and the full general ledger detail. So the scenario is basically we're finalizing a month end close, it is a month over month comparative balance sheet and P and L, and then two months of general ledger data. In the past, when I would give this to the previous best model, O3, it sucked at reviewing this. And part of the problem was a balance sheet and a P and L and a general ledger. That's a lot of information. And the AI model had a hard time knowing what to prioritize, which in my opinion, as a big brain accountant, you start with the balance sheet, right? And then you go to the P and L and you only really dig into the GL where necessary. But if the GL is 30 pages and the balance sheet's one page and the P and L is two pages, and I give all that stuff to the AI tool, it has a hard time not just getting tripped up on all little fiddly things in the gl. And so all I'm doing is I'm giving it these three financials and saying, like, you're an expert accountant, you're going to review this for factual correctness. I don't care about insights and about the health of the business. Just make sure the accounting is correct. And as an example, the very best that it could do before. And this is actually using O3 Pro, which was an even better version of the previous best model that was only available on the $200 a month plan. I'll read through the items that it found in a set of financials from basically the demo file in QuickBooks Online, Craig's design and Landscaping. I'll tell you the results from the Previous Best Model, O3 Pro, and in that prompt I tell it, order your findings, starting with the thing that is most likely wr and then, you know, kind of order them in decreasing likelihood of wrongness. So the very first thing or what it thinks looks most wrong, I want it to put at the top. And then as we go down the list, it's less and less likely they're wrong. So here's what the previous best model found. Number one, net income line on the balance sheet. It's probably wrong. It says because it doesn't match the May P and L figure. Well, the net income line on the balance sheet, it would only match the monthly P L figure if that was January. Right. So the balance sheet showing year to date net income, the comparative P and L for May and April, it's just showing those two months standalone. There's not a year to date column on the P and L. So the number one most likely thing that O3 found that is probably wrong is like, you can throw that out. It's totally wrong. Number two thing is that miscellaneous expense has $2900 in it. It says underlining bills are for normal costs, like $250 for hardware, $2,000 for insurance, 600 bucks for subcontract masonry work. That's a good one. That's a great point. Number three, pest control services revenue is negative. That's fine. Number four, board of equalization payable has a negative balance. Fine. Number five, funds has gone up. Most of these are like, yeah, probably. But then number seven is a large negative opening balance equity. And so on the balance sheet, it's negative 9905 in May, but then positive 5000 the previous month. For me, that's like, that's probably number one, either one, two, or three. And like, the magnitude of, well, that's very obviously wrong. Opening balance equity shouldn't have swung 14 grand in a month. So a few Other items on the list, they've got nine items in total that are like, yeah, those are probably good. That was the findings of O3 Pro, the previous best model prior to GPT5. Here are the findings. Findings of GPT5 for one, it blocks them out into two sections. First a section for balance sheet issues. Second, a section for income statement issues. We'll start with the balance sheet. One issue number one that it finds a 200 transfer appears one sided. So what it did is it went into the general ledger detail, it found a transfer to a savings account. But then on the checking account side, that $200 transfer was batched with several other things that came out of the checking account account. And so on the split account and the GL report it just says split. It doesn't detail out the $200 that went to the savings account. It says this is probably wrong, not looking good, right? That's not wrong. That's just an issue with the reporting. In fact, I don't Even know in QuickBooks Online how to get a general ledger report that doesn't like present each item. For example, if, if you pay 800 bucks out of the checking account and it goes four different places, I don't know how to get that to show up as four lines under checking on the gl. I think you have to actually use the journal report. Anyways, this first one it finds it's not an issue, so that's too bad. Second thing it finds opening balance equity. Okay, that would have been high on my list too. This episode is brought to you in part by Stanford Tax, one of my favorite new tax like AI tax workflow tools. Stanford Tax actually heard from a big number of you very happy users. Stanford continues to evolve and one of the big shifts in 2025 is that is it is now not just a tool for like, like serving as a portal for your clients to submit information into, but it's also just an internal tool for your firm now because it does so many useful things with those source docs. So basically you pile in all that info that the client gave you and it will compare those items against the pro forma checklist that it generated and tell you, okay, what do we have now? What do we still need? But it will also bookmark those documents, it will rename the documents, it will warn you about stuff like it being the wrong tax sheet. It will split bundled documents so if they give it to you in a single PDF, it'll figure out how that ought to be broken out. The number of years I was putting, like divider pages in and sitting in front of that scanner, if you know, you know. And Stanford now rolls all that up into a binder in the cloud where you can do your PDF annotations right there inside of Stanford Tax. Okay. Down with Adobe Acrobat. The world doesn't need us drawing more red squares in Adobe Acrobat. Okay. This excites me because it's pulling all this stuff into the cloud where every new AI model makes this whole flow better. That's super exciting. To learn more about Stanford Tax, check out the link down in the. In the, in the. In the show notes. This episode is sponsored in part by Ignition. Come, come, come, come, come. Here's the thing. Here's the thing about client billing. You can't bill for the time you spend. Billing. You can't. You can't. It is pure waste. And we're bad at billing because it's, it's icky. You let things slide. You don't get paid. You waste time. Waste time chasing people for money. That's not what I did. I would just. I didn't want to waste time chasing them, and so I would just. I just eat it. Don't do that. This is where Ignition comes in. The leading revenue generation platform for accounts and professional services. When I was running my firm, Ignition wiped out our ar. It's true. And significantly reduced the time spent billing clients, saving our partners over a thousand hours in the first year because they didn't have to bill all that stuff manually. Man, with the condition. We won more business. Business with impressive online proposals. No more hustling hustle PDFs and Word docs. We automatically pulled payment up front or at the end of a job when it was done. No more dialing for dollars. We transitioned our firm to more recurring revenue and better managed scope creep. No more leaving money on the table. Ignition is making thousands of firms more money and genuinely made us more profitable. Check out why over 7,000 of the smartest businesses run on ignition@ignitionapp.com. put a link in the show notes. Third one, there's a fixed asset, a truck at 13,495 bucks, but no accumulated appreciation. All right, fourth issue undeposited funds has gone up. But importantly Here it identifies two payments from the GL that ought to have cleared a 1675 payment and a $387 payment where it's like, these look too old. Specifically, look into these payments. That's a much higher quality answer than we got from the previous best model. It called out The Board of Equalization Payable that was negative but also said there's another current liability account called Arizona Department of Revenue Payable says consider making the tax agency naming convention more clear and consistent between these two accounts. Kind of interesting. Then it called out two loans that were new loans. A $4,000 one and a $25,000 one on the income statement items. Biggest issue was the miscellaneous expense 2916. Then it has sub bullets for every item in there and where it should be classified and offers to give you a reclassifying journal entry to fix it. And then it calls out some other stuff like in this, this accounting file there's like cogs accounts that are pretty much named identically and it goes through and it's like your naming here sucks. For example, you've got an account that is cogs colon job expenses and you've got an account that's job expenses colon job materials. It gives you suggestions for how to make that more clear. It even called out that there's an account lumber that is under the parent account Services income. It says move this to be under the parent account Sales of product income. It found an equipment rental that was under operating expenses that should have been under cost of goods sold because it referenced a job. And check this out. Number six meals and entertainment $28.49. It says if you track tax deductible meals separately eg meals 50% split this from any entertainment or non deductible categories. Good bot, right? That's pretty good. So here's the thing. The the GPT5 answer here it was very good with the exception of the very first thing that it returned with this. It's saying this transfer appears to be one sided. Now we could probably improve the prompt here to say assume like system things are working totally fine. Like QuickBooks wouldn't let you post a one sided entry, right? So you can probably improve the prompt here. That's like we trust QuickBooks to actually present this correctly, which maybe we shouldn't. But something like a one sided transaction couldn't happen otherwise. What it presented here, this is much much better than we've seen from any other model. And if there's one thing that this is tremendously useful for right now, it is the juniors in your firm. You think back to the early days of you being in an accounting firm firm and you just got so many dumb ignorant questions and how long it takes you to grow out of that. Imagine having this, the most patient tutor of all time where there's no dumb questions, unlike asking your senior a question and then seven minutes later having another question and you're like, oh, now I gotta wait a couple hours and look like an idiot, like asking him another question, right? I think stuff like ChatGPT, like AI, I think this vastly accelerates the growth of everyone entering the profession in a really exciting way. And it's funny how like us, us old farts, I say this as a 37 year old, are like, how am I going to make the time to use AI? And then you have kids coming out of school right now that have used AI to write every essay in the last three years and they're just freaking wizards with it. And they can't imagine entering the workforce and not working with AI every moment of the day. We're like, how will I ever make time to use it? While they're like, this is the only way I know how to operate. And we look at them and we're like, well that's problematic. And they look at us and they're like, why are you doing all this crap manually? And the right answer is probably some marriage of those two things, right? But those kids, kids coming out of school into the workforce, are they zoomers? I don't even know what they are. I think they're going to have such a faster growth trajectory with really good AI like at their fingertips. That's really exciting to me. So that first test, just a totally open ended review of a balance sheet income statement in gl, I would say it's a pretty big step up and again we're not prompting it in a specific way or saying here is how an expert would review it. We're leaving it up to itself to figure out how to review review it. Now there's a GPT5 Pro also which is only available on the 200amonth plan. And you better believe I have the 200amonth plan because I'm a nerd like that. I put the same task to GPT Pro or GPT 5 Pro. It worked a couple minutes longer and the answer was marginally better. But I wouldn't say it's like worth springing for it. In fact, the bigger issue with the Pro plan of Chat GPT is it doesn't come with the same data protections as the team plan. So it's not really useful helpful for business purposes. But just out of curiosity, I tested the Pro plan. It was a little bit better, but not a huge amount. Okay, second test. This is when we started doing, when that agent mode for Chat GPT came out specifically around Hustle and spreadsheet stuff. So what I've got is seven tax source docs. So W2's 1099, Rs 1099 interest, a brokerage statement, a Social Security statement, just government forms, 1040 source docs. But then I've got a Excel workbook from excel1040.com this amazing workbook this guy puts together every year. It's got like 60 tabs tabs and all of these input tabs for all of the tax data and then other tabs that recreate government forms based on what you put in the input tabs. So I gave this to GPT5, the spreadsheet, and then the seven tax documents in PDF form. And I say take all the data off these PDFs and put it into the Excel workbook. Now, chat GPT agent, when you enable that agent mode, it does this flawlessly right now. Takes about 10 or 15 minutes for it to do it, but it will pull several hundred figures from the PDFs and put them into the correct places in this 60 tab Excel workbook. It is pretty wild. Now there's got to be an upper ceiling to this. Like I'm giving it seven tax docs, but what happens when you give it 50 tax docs? I suspect it won't work, but it's not because the AI model isn't good enough right now. It's because it isn't allowed to work long enough yet. Which is an interesting distinction, right? Like that just means that on my 30amonth chat GPT team plan, that's going to burn down too many servers for me to be allowed to have a thing that can work for like an hour at a time, right? And so that's actually a new feature that I need is the ability for it to work longer. And AI models are actually good enough for that. But anyways, ChatGPT agent, it does this flawlessly. But I gave this task to GPT5 and I couldn't get it to work. In fact, I couldn't even get it to make a functional Excel file. And I'm giving it the input spreadsheet. I'm like, this is like the workpaper template we use in my firm. So just modify this. I tried this several times, couldn't get it to actually complete. And the couple times it said it did, the Excel file it gave back wouldn't work work. I then tried GPT5 Pro, the one that's available on the 200amonth plan. I'm like, maybe it just needs more horsepower or something. And Bless its heart. It's been thinking for about half a day now. I'm pretty sure something's broken, but that doesn't feel very usable to me. So, for completing an Excel spreadsheet thing, I still think agent mode is better. So instead of just using GPT5, hit the little plus button right next to the field where you type in your prompt, hit the little plus button, toggle on Agent mode, and if you needed to do something with a spreadsheet, use that agent mode instead of just GPT5, which is in my experience, it's just kind of been a mixed bag with spreadsheet stuff. But agent mode, very, very good at spreadsheet stuff. All right, last test I did number three, giving it a checklist of review steps and the financial statements. So test number one that we did was not giving it any checklist items, just saying, review this stuff however you want to do it. But this test is reviewing it by performing a specific set of steps. And I'll be honest, I'm not sure if the future looks like like just letting AI review it however it wants to review it, or if the future looks like it reviewing it according to our prescribed process. Right now, the latter feels easier, right? Because we can trust that it's actually checked the things that we want it to check, rather than it just choosing its own adventure. And this is something that historically, models haven't done very well. So a very basic test that I've used back several years now is getting it to review three things on a balance sheet. Here are the three things. You ready? First, make sure it's cash basic. Second, make sure total assets equal total liabilities plus equity. And third, make sure it is a comparative report, that is it is a current period compared to a previous period. This is the format that we sent our financial statements out in my firm. There's a hundred other things that ought to be on that checklist, but let's start with three. And to date, AI hasn't been able to do those three things. And so, for example, if you throw this at the 4o model, it gets super tripped up on, for example, QBO, even on cash basis reports. Reports will put accounts receivable or accounts payable with zero dollar balances, even though it's a cash basis report. And four, zero will come back with things like, well, I see AR here, so it's accrual basis, or oftentimes it will say, oh, I see a credit card balance here. So obviously it's accrual basis. And when you see that it Just makes you sad because you're like this thing's right, this thing is like top, you know, 99.9 percentile in the world at writing code. But it sees a credit card payable and it thinks that makes it an accrual basis balance sheet. Like, like that's, that's annoying. So 4O didn't do this well. The previous best model, O3, it did a better but still not great. It would still make dumb mistakes like that. In fact, that was where I saw the credit card payable example was with the previous O3 model which is just wild but ran the same test with GPT5. I'll just tell you the findings is where on the previous models I had to do a whole bunch of detailed prompting like telling it, okay, if, if you're going to decide whether it's cash or accrual basis, like usually somewhere on the financial statement it will say that and look for that and to determine if it's a comparative financial like you'll one month in the previous month or one year in the previous year. And I had to like tell it how to do all this stuff and it would still do it wrong. With GPT5, all I did was gave it a very concise prompt. Number one, figure out if it's cash basis or accrual basis. Number two, make sure total liabilities and equity equal total assets. Number three, make sure it's a comparative financial statement. That is it compares a current period to a previous period. That's it, no other instructions. And it nailed it. Tried it on a few different balance sheets, totally fine. Now that's not rocket surgery, right? Like that those three things, that's not super hard. But this, this is kind of like the tax work paper example where once the model's good enough to make those determinations, the bigger question then is how long is it allowed to work? Because having it go through a hundred item checklist on every single set of financial statements that we do, that's pretty interesting, right? That's pretty juicy. The day we get to where AI can complete a checklist as good as a human, which by the way not a high bar because as soon as a human sees a checklist like their brain goes into this sort of autopilot where they can't check those things faster enough. But if I could put that same checklist in front of an AI, I mean I wouldn't put 100 checklist, 100 item checklist in front of a human, but in front of a computer, no problem. In fact, that's that's kind of exciting, right? The notion that we could put endless checklists in front of this to make sure we never make that same mistake again. And so if it can do three at a time, what we actually need is the ability for it to think longer. So it could also do 300 at a time because if it's good enough to do these three things, it's probably good enough enough to do a little more complicated things just in a more modular way or with more time to think about it. So I gave it a slightly longer checklist. I added three more things. I added a fourth task for it. I said make sure that there's no accounts presented that have zero dollar balances for both periods. I was trying to go back to like my. When we do compilations and reviews and stuff like that, like what are the formatting things we would look at that were kind of the policies in the firm for how we would present financials? Fifth thing was, was make sure that we're always using dashes instead of zeros. That was a formatting thing we did on my firm. And then sixth thing, identify any accounts with non normal balances. That is if the normal balance is a debit, then call out if it's showing a credit or a negative balance. So I checked this at GPT5, did it flawlessly. Really good. Six things called everything out exactly right. I don't know what the ceiling for this is, how many items it can do. Whenever you're trying to get AI to do something for you, you kind of have to think of what is the cumulative cognitive load load of what I'm asking for. Six really easy things, maybe it can knock that out, no problem. Six really big complex things. I see the folks that just don't get it with AI yet. This is usually what they're doing is they're giving this 30 step process to AI and then it falls on its face and it's like they're like AI is dumb. It's like no, like you gotta chunk this stuff out. For the very practical reason that OpenAI right now won't let chat GPT work for an infinite amount of time because that's very expensive for them and I'm only paying $30 a month month for this. So I always think about like how much am I asking it to do at a time. But these six sort of fairly simple review tasks it had no problem with. But I think the more interesting product of this test is the last thing that I have it do, which is create a work paper to show your work right. We always gotta be thinking about AI almost as the intern in your firm where it's like yes, it can do really cool things, it's very useful. But the provability of what it did, that still matters too. Because I'm not going to blindly trust the output. I then have to prove. Prove is like what it did actually correct. And if it takes me twice as long to do that, that's a pain. So I on both the three checklist item and the six checklist item prompts, I had it then create a PDF work paper. Page one of the PDF work paper was supposed to summarize its findings for each of the checklists or the the checklist Items. So page one has you know, 1, 2, 3. In the case of the six, the six item one for all six, it would identify its sort of summary findings on page one. But then I told it after that include the balance sheet and annotate it in a color coded way to make it easy to go from page one and see what on the balance sheet after that you saw that made you answer it that way. And the previous best models, it would really struggle with this work paper prep stuff. But it was adorable to see it try. Like it would draw like little red circles and squares and arrows. It was really, I mean it was almost like my children trying to mark up a PDF in a very rudimentary way. It was cute to see it try, but in the end it wasn't all that helpful. Helpful now GPT5 work paper prep. Big upgrade buddy. Today's episode is brought to you in part by Team up. Who helps you hire talented accountants in the Philippines. Did you know you can just hire someone in the Philippines? That's the thing that you can do. Similar to hiring cross country. Like most of us run distributed teams these days, you can hire somebody that's in the Philippines. Now oftentimes big question if you haven't done this before is, is security? What about security? To which I'd say what do you do for security for your onshore team? If I hire somebody in North Carolina or for goodness sakes Florida, what do we do for them? Well, probably ought to give them a specific work device. Have like on device encryption, the ability to disable the machine remotely. All of those things, important things to do for any hire. Exact same things you do when you hire somebody in the Philippines. Only real difference for US firms is you have 7216 disclosures. If you do tax work, actually if this person will do tax work and for Most firms that is, that's two disclosures that you got to put in front of the client. Not a big deal. Team up, they've let loads or they've worked with loads of firms that have navigated all this can give you the verbiage recommendations. I did it myself and my firm disclosures and everything wasn't as big of a thing as I built it up to be. And I went 4 for 4 on my Philippines hiring and I had never gone 4 for 4 with entrepreneurs hiring. Maybe I got lucky. I didn't get lucky. There's like a different attitude towards work with folks in the Philippines. Literally three days ago, somebody at an event, we're talking about their team and they're like man, it's, it's like refreshing. Anyways, if it's something you've ever considered, chat with the folks at Team up as I think this person had, they actually hired those folks through Team up as well. It's one of those things that seems more spooky until you actually do it. Learn more about Team Up Check out the link down in the show notes. This episode is brought to you in part by Client Hub, the Practice Management System. If you are considering a new practice management system system, there's no way that you're not overwhelmed. Let me give you the TLDR on who are the people that need to be exploring Client Hub. Specifically Client Hub started out as a really good sort of two sided task management system for bookkeeping firms. So you've got your own tasks to manage. But we also push tasks to clients when they're holding us up and we need information from them to get that stuff done right. Cloud was one of the kind of the client portal first PMs that makes that as as smooth as possible for your client clients. But increasingly they're getting into all things workflow, specifically your bookkeeping workflows. And so if there's one thing it does really well, it is for folks that don't want to have a standalone like month end close automation platform, but then a totally separate PM over here, Client Hub might be the best marriage of those two things in a single place. So if you want more like power user month end close automation stuff, but a version of that that is like inside of your practice management system system worth giving Client Hub a look, check out the link down in the show notes. So on the three checklist item test, the PDF that it comes back with has a very nicely formatted page one and each of those items, item one, item two and item three that like the heading for each of those sections is shaded in a specific color. And then you go to the balance sheet after that. For example, like the first item that's figuring out is it a cash basis or accrual basis balance sheet that is shaded yellow on page one. 1. You go to the balance sheet after that. Wouldn't you know it, the little cash basis label on the balance sheet is shaded in blue. Did I say it was blue on page one? Whatever, it got the color right. You go to number two. It nicely shaded the column headings that said, you know, March 31st and then April 30th side by side. It shaded that as proof that it was a comparative balance sheet. Number three was do total assets equal total liabilities plus Equity. It did all the coloring on page one, but didn't actually mark up the balance balance sheet. So everything about the work paper looked very good and convincing and almost human. In fact, page one looked better than what a your average accountant would give you. The only thing it missed was the shading on the total assets and total liabilities and equity. And really that ought to just be like highlighting the total assets line and then highlighting the total liabilities and equity line. So I didn't get that, but it was a very big step forward from the last model. I tried to do the same on the six step, the six item checklist. And the work paper got worse. It still did the page one very good and had different colors for each of those items. But the markups on the balance sheet itself, it was a lot to look at. It was a little hard to read. And all this brings me back to so kind of, what are my three main thoughts that I'm left with here? Having done some initial testing and where do I see this going in the near term? First big thought is just the notion that we need to build out sequential steps, build out workflows in more modular ways ways. And like building out this work paper is a great example. GPT5 can do three checklist items in these basic examples very well. It can create a pretty good work paper. But I can't give it 30 items. And it would be really nice if either I could give it 30 items and it would chunk out the process itself. Or if I could build in more modular ways, like distinct little workflows that were bite sized enough for it to handle and then have it immediately hand that off to like the next workflow. Run that stuff in the back background. Now I think there's a version of this that could almost be like GPTs So GPTs or custom GPTs are like little assistants you can make in chat GPT to do a very specific job. So, for example, I can make a GPT for the HR documents in my firm. I give it the employee manual and any other knowledge it needs to have. I say all you do is parrot things from the documentation to answer questions, and then I can share that GPT with my team so they can just go to the bot, ask it a quick HR question, rather than have to find all the potential places that something related to that could be in our company documentation. Great time saver. GPT is a great way for you, maybe the biggest AI nerd in your firm, to create tools that can be used by the rest of your team. But I would love to have something similar that's more workflow oriented, where it's like, how do I run like these 5 GPTs in sequence? And the results of 1 GPT hands off to the next one? And they probably wouldn't be called GPTs, but like a more sort of workflow angle for handling larger tasks, tasks that we could chunk out. And there are some tools out there that will do this right now. One is called Lindy, one is called Gumloop. They're almost like Zapier if you've ever used that. But for AI workflows specifically, the problem is I just want that stuff inside of ChatGPT because I don't want to have to pipe all my client data through yet one more system. And so I'd love to see a way to more sequentially handle bigger, more complex workflows. Second thing that's, that's a takeaway for me is we Talked about how GPT5 is held out as like the most advanced coding model there is. Now, I still don't think that we think enough about AI as the tool to build the tool to do the job, if that makes sense. So take for a basic example Excel. Like macro scripts. Most of us don't know how to write macros, and there's a good number of us in the audience who do, and that's fine. But for those of us who don't AI right now, it can build like flawless scripts for your spreadsheets. I would encourage you give it the actual spreadsheet that you're trying to write the script for, and then it'll use the correct, like tab workbook names and stuff like that, but it'll write flawless macros for you so that any sort of transformation or recurring job we do in a spreadsheet, we can now use AI to write a macro to do this for us. But as you know, AI can also write like entire apps to do specific jobs. But the problem with its ability to write code is then where does that code go? Because if I have it write a useful little app for me and my team right now, it is very hard to put that app into, you know, production or into some environment where the rest of my team can use use it. This, I think, is another opportunity for like a. A sort of custom GPT style. Here's this little app and I can share this app with the other members of my team, because then your AI champions, your most competent AI people in your firm, they can make these little micro apps that your team can then use also. And it can all just live inside of ChatGPT, similar to how your custom GPTs do. Now, the whole notion that, like, AI gives you the superpowers to write code and sort of create apps now, like, that's wildly powerful. And the code writing is enough to do that right now. The problem is we're not like developers that are working on big code bases that are already in production. We're like, I just want you to draft up like this dashboard for me with a knob here and a dial there. And it could do that very, very well. But then where does it go? That's the problem. Doesn't have anywhere to go. And so we could totally scrap together all sorts of useful things for our own workflows. The problem is they're not shareable. They're not like, persistent as, as the AI models continue to get more powerful, I would argue right now they're plenty good for writing very good code. They have been for a while. Bigger problem is, what do I do with that app that I just made or generated the code for? I need somewhere to put it and to be able to share it with my team. Last number three. Third sort of big thought coming out of this. We're real close. In fact, we're already there in, in basic applications, we are real close to tools like models like GPT5 simply being able to function as an accounting ledger. So, like, I give it a bank statement, it classifies those things automatically and gives me a balance sheet and a P and nailed back. And that's like, unthinkable and may. It may almost feel wrong. It's like, how am I ever going to be able to trust that? Honestly, like, there's this old Paul Graham quote. I think it's Paul Graham. He's like, you've already seen the Future you just dismissed it as a toy. And so this, the notion of AI being able to do that, right? You're like, no, I have a hard time seeing it actually do that. And it would only ever be able to work in very, very simple applications, right? So let's just start there because right now I think it already does this. Good enough for simple use. Use cases, client gives you a spreadsheet or client gives you 12 months of statements that have like virtually no activity in them because they've got this standalone LLC or whatever this dumb little side business is that has virtually no, no volume. And you're like, I'm not, don't want to do the bookkeeping for that. It's not big enough to justify it. Like almost every entrepreneur has some of these sort of dumb side companies that are just like off gassing from their core, core business. And it grinds your gears to have to pay for another cloud accounting platform subscription just for this dumb little company, right? Feels like a waste of your having to do the books for this simple little thing. Well, what if we could just chuck those statements or whatever information was required into a chat GPT conversation and it produces for you a balance sheet P and L and a gl, maybe even gives you like an Excel download. It is already good enough to do that in very simple ways. And when there's so much debate right now around when are we ever going to get another accounting ledger that's like that will rival QuickBooks and Xero and when are we going to get the next AI accounting ledger might take is we're looking at it. If you compare all the horses in the race, I'd have a hard time betting against ChatGPT just straight up AI models doing this really well. And today you can only really see that in very simplistic use cases. Right. Very small companies. But as AI models continue to improve the capabilities and what it's able to do like that just keeps getting better and better, right? And this would already be tremendously useful for us as accounting pros. It ends up where we end up collecting all these very small, very simplistic accounting files. Maybe you've got an Excel workbook or a QuickBooks desktop top file where you're tracking this over time, man, wouldn't it be nice because these models can already flawlessly pull transactions from bank statements. Like, I mean we used to use, you know, auto entry and Dext and we used to use all these tools to be able to extract transactions from bank statements. That's the thing that it can Just do now, no problem. But to be able to just chuck whatever information that you have into it and it makes sense of all of it and then organizes it for you, man, that'd be nice. Nice. In fact, if we go to the tax source document example here where it pulls all the data off of it and it puts it into an Excel workbook for you, you can see how it's like, dang, that could be real useful to me real soon. In fact, 12 months from now, 24 months from now, when it can handle 100x the volume, you're like, wow, we're almost certainly using that. Like that's almost certainly saving me a pile of time. Right. So where does this live alongside our specialist tools, like let's just say the full blown account ledger. Well, I think it starts, that is chat, GPT starts at the bottom of the market and the question will be how far up does it ever come? Does it actually challenge, you know, a million dollar company with very complex needs and all that? I have a hard time imagining it getting up to that because then aspects of the user interface and user access, what people have access to, there's much more nuanced problems then, but you could certainly see it nibbling away at the bottom of the market in those very simple use cases. Pretty interesting and would be a tremendously useful tool for accounting firms where it's those little trivial things that don't have an off the shelf software solution that are our main headaches right now. Like I don't need another tool that will, you know, help me reconcile my stripe data back to the accounting file or some of the bigger, more mainstream problems that a lot of platforms are solving for. That's all fine and I want them to keep working on that stuff. What's left are the very nuanced headaches that I have today that no platform will ever solve for. Because this one client has this one specific problem and it takes us three hours a month to, to fix it. That's where AI assistants like this are tremendously powerful for us. That's where I may use the machine to build a machine that is AI to write code that is like a custom little bit of software or Excel macro that'll do that three hour job for me each month. You know, I always hear and like glorify these people where they get into this job and it's just like jockeying spreadsheets all day and they find a way to automate it and it turns out they don't actually work at all. Man, you better believe like that stuff's happening at record rates right now with AI assistance assistance. And that's, that's great. The less time we have to spend on those dumb, trivial, like client specific things, the better. So all in all, GPT5, it's been a frustrating launch, but I'm seeing promising things right now in terms of how this changes our behavior. I don't think it actually changes our behavior inside of accounting firms at all. The ceiling for what it can do is now higher. But the day to day things that we do as a result, they ought to be the same. We ought to have chat GPT up on one of your monitors staring back at you so you remember to use it. We ought to be meeting with our team every single one week, sharing, hey, what did we try this week? What worked and what didn't work? Like just creating that space for people to come together every single week. I think right now is essential because just by being near each other, we get that contact high of seeing what other, seeing what other people are doing and opening our minds a bit to how we could use it. And honestly, that's it. Like get the tool in people's hands, specifically chat GPT team, but then get everybody together each week and you wouldn't believe how many problems that solves from like, well, how do we keep up with each new model? Well, if you're meeting every single week, it'll be part of the conversation. Hopefully on your team you've got some folks that are like excited by this stuff and it comes easier to them and they can bring a lot of the energy to those meetings so you don't have to do it all yourself. Right? So GPT5, very exciting stuff. Definitely push the bar higher for what's possible. But the day to day routines and how we're leaning into AI, I don't think it's changed at all. We got to keep doing the same things, we got to keep leaning into the same weekly hacks, habits and hopefully now AI will take us even further. It'll automate more of that stuff that we don't want to fuss with. Right? Hey, if you're new here, subscribe. Man, we talk all things accounting from running oftentimes AI. I think the folks listening to this podcast are like the most ahead of the curve folks when it comes to accounting AI in the profession. And don't you want that to be you? Right? Any cool GPT5 stuff you've done, drop that stuff down in the comments. If you're listening to the audio version in the show notes. You'll see a link to the YouTube version. Come join the conversation there. Drop any testing you've done down in the comments. We can share some questions, cool ideas. I'll see you in the next one.