I Ran 5 Accounting FIrm Tasks Through Claude and ChatGPT. Here's The Winner.

Hey, this is Isaac from Ossa.co, an automation agency that has worked with over 100 accounting firms. And you've probably heard people raving about Claude, leaving ChatGPT behind. But the question is, why? And does it actually make a difference for the work that our accounting firms perform on a daily basis? To answer that, I'll be actually testing and running both Claude GPT and Claude's latest models through 5 accounting firm use cases to see if there's a real difference. With that out of the way, let's get into it. So we'll be testing the models through a general ledger review, trying to find a discrepancy, allocating payroll and creating a general entry, building a financial model, and then building a dashboard from a client with their financial. Let's start with the first test. So I've taken a file already with some transactions, very simplified general ledger with just a couple of dozen transactions, and we've inserted several issues to see if Both AIs catch it. If we take a look at Claude, it goes ahead and reads them. Here's what the prompt says: Here's a monthly PNL. Review every line item and flag anything unusual, inconsistent, or worth a second look before month-end close. For each flag, explain your reasoning. So it went ahead and found the literally the exact five issues we inserted here, notifying that you know the rent was higher than expected. There was a professional fee that didn't really have a memo. There was a specific expense that was a meal, but that it was categorized as equipment. There was a duplicate, and there was like a big revenue drop. That is, I think, worth questioning. One of the things I really liked about the rationale Claude provided is that it also kind of went above and beyond just the financials and said, hey, there might be a tax benefit here of like categorizing this. So this might have a tax implication for the client. I also like the conciseness of this compared to what you'll see here with ChatGPT. It literally went line by line on the P&L to try to assess what's wrong or what's not. Although I understand that I texted technically told it like review every line item, I was not really expecting it to literally create a table with every line item. So I think that it kind of like took the instructions too literally. Just something I've noticed that you have to be careful with ChatGPT in particular. Claude seems to better understand the intent behind what you're telling it and specifically what you wanted to see as the output at the end of the day. Although to be fair, ultimately ChatGPT also found all the issues that we planted in this particular example. For reasons described though, I'll probably give this particular test— we'll put Claude as the winner of this one. Let's take a look at the second test, which is essentially having a bank statement and a general ledger cash report and seeing if it catches again 2 issues that we've inserted intentionally. So we've given it one that is an actual kind of recording error that needs to be fixed, but there's another one. This is just like a check that was— that's still outstanding and that was released towards the end of the month, and we wanted to see if it was gonna be able to catch that and actually identify that that particular outstanding check doesn't really— even though it's the discrepancy, it's not an actual error, but just a timing difference. If we take a look at Claude, here's what the prompt looks like. Here's a bank statement and a list of recorded transactions. Identify any discrepancies, flag unmatched items, and format the output as a reconciliation work paper that a reviewer could sign off on. So very simple, just 3 lines. It went ahead and see what ties, what doesn't, what's the difference, then try to understand the difference behind it, and ultimately came out and said, cool, we catch the outstanding check is not yet cleared by the bank. This is just a timing difference, as well as suggesting here there's a recording where there was a $90 difference here that actually needs to be corrected. If we look at the Excel, I like to also get a sense of like Excel usability. It created 5 different tabs, 2 of them which were already there, the GL account and the bank statements In particular, but it created the reconciling items of like, hey, these are the two things that we need to look at. The age of those things, the status. It's just saying, hey, this just needs monitor and no real action required once the clear happens. Here's all the transactions to see at a very detailed level where it found the mismatch. And here's suggesting, hey, if those two things are fixed, we should be good to go. So pretty usable Excel to kind of like document its thought process. And you being able to here come in and just give it, and you being able to come here and just tell it, you're you're good to go. If we take a look at ChatGPT, it follows a very similar process. It essentially identified the same issues, so there wasn't really a big difference. The only thing I'm noticing here, for some reason, this keeps coming up. I think it might be a bug. I don't know if it's only on my end. That's a very small pet peeve there. I do think the, you know, it did the same thing, like all the details to assess where specifically are the mismatches, and here it kind of listed the exceptions. Noting specifically that here there's not an adjustment that's needed. It's just a timing issue, and you just have to confirm it is. It clears in April. If we take a look at the reviewer notes, it's just effectively reiterating that a little bit more in a different format. So I do find it. I do find it somewhat repetitive. It does feel like this is more just an additional tab with a summary of the things it found, which in this case Claude didn't really provide because it perhaps felt like. This line item detail, this action items, and this like reconciliation summary was more than enough. In this particular case, both created very usable work papers and Excels. Both found specifically the issues that we planted, including the fact that the check wasn't really an error that needed an adjustment, but simply time to elapse for that to be cleared. So I will give this a tie because I think both models performed really well. As a third use case, we're going to have a payroll register report with 6 employees across 3 departments. I need to book payroll were journaling for that particular company. So here's what it looks like on Claude. If we go to Claude, also worth noting, I told it to specifically put it in a standard QBO format. I wanted to see if by me just using like acronyms and just very language that you probably use with another team member, if it would capture that and actually create it and know what that actually— so it just says, here's a payroll register. Create a journal entry to record payroll, allocating wages by department based on the percentages shown. Include employer taxes and net pay. Let me briefly show what that particular report looks like. So, here you can see the various columns, and although in reality payroll reports would look different from different providers, this is just like a dummy data example to see how the models perform. One of the things I didn't like about Claude, even though it gave me the details, it showed me a table, it gave me some of the rationale behind how those numbers were built, it didn't actually build Excel or CSV document for me. I have to like prompt it to do that after that. But of course, after Providing that it did pretty well, making sure it has like a journal number, journal date, account, debits, credits, description, name, and class, which it was able to incorporate there as well. Notice that it even had like some sort of naming convention for the journal number, which I didn't really ask it to do. So that was a good extra there it did for me. If we take a look at ChatGPT, very similarly, I would say the only difference here is that it was able to understand from the very beginning what was the QuickBooks-style journal entry workbook that it needed. It understood that I wanted it to just be able to Although it's really inserted this, this total, which actually doesn't really allow me to directly insert it. I will still have to like download this, delete that to download it and push it to the system. Also gave a little bit more detail on the description, but I'd say this is pretty comparable to what Claude did as well. Just because the detail of not being able to capture the QuickBooks format off the bat, although both did really well, I'll give this particular win To ChatGPT. Next, I want to test the model to create a financial model. I'm gonna give it some actuals and some assumptions, but I do want to get it to test creating 12-month forward projection in a table format: revenue, cost of goods sold, gross profit, net income. And not only follow my assumptions, but see is there anything else it can infer or some sort of adjustment it can make to the model that may add value. So let's take a quick look at what that looks like within a Claude first. So a 12-month projection, it says here's a 12 months of financial financial data and a set of assumptions. Build a 12-month forward projection in a table format showing revenue, cost of goods sold, gross profit, operating expenses, and net income by month. Follow the assumptions closely, but make any reasonable emphasis or realistic adjustments. It went through, verified no errors across 143 formulas. Despite this being a relatively simple, simple model, it showed it in a table first, as we saw earlier. Sometimes it likes doing that, but it did come up with the Excel I requested, and it told me, hey, I've added a new tab, which is 2026 Projections. If we open up that particular Excel, we can kind of verify that the formulas are working correctly. There are even comments on it just to kind of clearly label what these things are. If I were to change this to 12, then I should see sort of the revenue adjust. If I look at cost of goods sold and gross profit and I change this to 55, then I should see that adjust as well. And lastly, if I put a higher year-over-year growth on the operating expenses, then I should see a lower net income. And I see that it also like added a net margin column as well for me to see the percentage. That's actually the only thing it decided to add. It said, hey, your assumptions, that looks good. Here you go. So, it says the one reasonable addition I made was net margin for easy reference, which I think is pretty— I would have hoped it kind of did more than that. I didn't ask it to do more than that, but that's kind of the additional adjustment or change to the model that it came up with above and beyond what I told it to do. If we take a look at the model that ChatGPT created in this case. Let's take a quick look at this one. Very similar, very similarly structured. It has some of the line items. It kind of separated it a little bit differently, but fundamentally also has— what is this inconsistent formula? But it still has like a basis at the source of each one. If I go ahead and change this, I should be able to see this impact the various numbers. I do like that it even has sort of like this red formatting within it so that effectively if I increase this a ton, I should be able to see a lot more red. So I'd like to say that, you know, both did pretty pretty well on this particular test. So, that is again a tie from my perspective. All right, so as the last example here, I'm gonna give Claude and ChatGPT a simple annual P&L dashboard. I'm gonna give it a P&L Excel. I wanted to create a dashboard as well as create an email that I want to send to this client to tell a little bit about the story of his business based on those financials. If we take a look at Claude on this test 5 and last one, I just tell, hey, here's your full-year P&L monthly breakdowns, create a client-facing financial dashboard that a non-accountant business owner could understand, highlight key trends, flag any concerns. I code it so that I can make it public. So you can see here the final HTML that it was able to come with here that I can now share this particular artifact. And it's saying, hey, here's a clear, jargon-free look at how the business performed. You finished this year breakeven. I broke a breakeven. Here's a small loss. A tough sprint pulled you into the red, but a strong second half nearly erased it and left you with real momentum. I like that it highlights the good, highlights the bad. It had some like coloring, some charts that make a lot of sense in terms of like how, uh, where the money has gone, uh, the simple split, what to look for. So it even came up with suggestions as to what to look for and what to watch for. I do like the, the writing, the style, the design of it. It's a lot, uh, very digestible. I also like that Claude often comes with a couple of versions, uh, and emails. In this case, warm and personal and brief and professional. I'll let you read it to make your assessment. I read through both of, both of them. I kind of like when it is bit more brief. Sometimes these types of longer messages feel very marketing-like, like, no pun intended, just what the numbers are actually telling you. Like, I don't know if that's something I would actually write in an email myself, but this one felt pretty good and something that would feel comfortable sending on my behalf. If we take a look at ChatGPT now and look at the client financial dashboard, we should be able to see the preview here. And this is what it looks like, a little bit more bare bones, though, so very like vibe-coded, if you like. And No real headline. So that plain English view of a company's full year profit and loss is kind of not really directing it to the person. And it's having details that don't really matter, you know, like static HTML, no database, who cares, right? Like that's not something I want to put in front of a client. So it kind of like really didn't omit those details that didn't matter. Revenue fell sharply does kind of point out some specific things. Marketing spend needs ROI visibility, kind of highlighted that. A chart that, you know, makes sense. It's not necessarily like as my, from my perspective, not as user-friendly as the one that Claude created. It, but ultimately, it kind of got the job done. It also has the email here, just one version. I read through it, I find it pretty long. Perhaps it's just a kind of function of a better prompt, but I did like that clause, so able to come up with a couple of options. So, in this particular case, better story, more engaging design, graphics that are better, easier to digest, 2 email versions excluding one that's very brief, and I think more applicable to what a human would want to send. So, this one goes to Claude. Now, to wrap things up, I think like ChatGPT in the last several weeks has really caught up working with documents and Excel. I just feel like still lacking in writing and design. Claude is incredibly good at the visual components, as you saw. It writes better stories, and it feels like more human, you know, like it has a better sense of what actually matters. It requires less prompting and understands the intent behind what you're telling it, and it doesn't take things as literally as sometimes ChatGPT In this particular example, so from my perspective, the Claude versus ChatGPT debate doesn't matter. It does in the short term, and the long term is a distraction from a more important question, which is like: What are you actually using AI for in your firm? Ultimately, the firms that use either, I think, were better off than the firms that use none. So the most important thing is that you use AI consistently in your workflows with prompts that improve over time. If you understand how firms are actually using AI to save time, feel free to check our audit offering in the card that will come up right now or in the link in the. Description. This is a process that we do to consult with firms and their teams to identify and estimate what are some of the biggest efficiency opportunities that they have on their business. Other than that, I'll see you on the next one, and until then, stay efficient.