Frontier AI Models Outperform CPAs on Month-End Close Tasks at 1/10th Cost

[music] [music] Hello and welcome back to the accounting podcast, your weekly roundup of news in the profession. I'm Blake Oliver.

And I'm David Liry. and David. This week we're talking about a new study that found that Frontier AI models perfectly completed month-end closed tasks, outperforming every CPA, 12 of them, at less than onetenth the cost.

We're also going to talk

so much in that that that study that I can't wait till we discuss it. It might be half the show. That's how much is in that study. It's amazing.

And then you've got a story about uh what clients are willing to pay for accounting services.

Yes. which is still a decent amount and it's interesting how it changes as they make more revenue, but it's not like linear. There's a there's a big shift that happens at like a million in revenue. So, we're going to dig into that and as much accounting news as we can after we thank our sponsors.

So, our sponsors this week, we have Cloud Accountant Staffing, OnPay, Finance Accounting Technology Expo, and Savant Labs.

And I was just going to read that ad for you, David, but I got signed out at Google. Oh,

so I need you to read the first one.

Yeah. Are you tired of the endless search for qualified accounting talent? You're not alone. Growing accounting firms are struggling to find available and affordable team members when they need them most. Cloud accountant staffing has a solution with the revolutionary candidate portal. Unlike traditional staffing agencies that waste your time with sales calls, paperwork, deposits, the cloud accountant staffing candidate portal gives you instant access to highly vetted qualified accounting professionals. No waiting, no hassle, just top talent. right now. What makes this difference? Speed and simplicity. While other firms make you wait weeks or months, with cloud accountant staffing, you could interview someone as soon as tomorrow. Their boutique support ensures you're actually getting quality talent that's both available and affordable. Exactly what a growing firm needs. The candidate portal puts you in control. Browse live candidates, make selections on your timeline, and build your offshore team without the traditional headaches. To find, review, and book interviews with potential team members all in less than 10 minutes, head over to accountingpodcast.promocass. That is a cloud accounting or I'm sorry, that is accountingpodcast.promocas. Thank you. Three coffee emojis made a live event. Happy federal fiscal new year and happy end of 2026 extension season. Ashley says, "Happy Friday." Edgar, what up accountants? Uh Shelley, happy day after deadline to those who celebrate. Great to have you with us. Uh and if you haven't checked us out on YouTube or seen a live stream, go to YouTube, search for the accounting podcast, hit subscribe and that notification button icon and you will uh get to see us live if you can catch a live stream. All right, let's talk about this study. I feel like, well, I personally am not surprised by this because the recent uh um new the newest models like Opus 5.5, Fable before that have just been incredible. And I'm using them to do all sorts of work that I never could have used before. And now we've got a a study by a startup that specializes in AI for professional services that has demonstrated just how capable these models are. That company is called Merkor MC and they are a startup founded in 2023 that specializes in using professionals working with professionals like CPAs to train models. So teaching the models, evaluating them, giving them feedback so they can get better. And if you think about it, that's the only way that we're going to get models that can do the work that CPAs can do. And I think that's in big part why uh the recent models have been so good. So this company did a study and it's called human baselines for benchmarks. AI now outperforms junior accountants. And basically the takeaway is that the latest models have perfectly completed month-end closed tasks. They can do them perfectly 100% of the time. And they outperformed every CPA in Murker's study at less than onetenth the cost. What did they do? And when I say onetenth the cost, we're talking like 21 cents a task. Like crazy crazy cost reduction we're seeing.

Yes. They tested 12 accountants on simplified tasks from their Apex accounting benchmark and they found that Frontier AI models were faster and more accurate than every participant. The accountants were described as junior but they had 5.5 years of experience on average and all of them were licensed CPAs. The participants completed four realistic month-end close scenarios. searching company working files, identifying the correct numbers, performing calculations, and delivering tables of results. The study was designed to measure AI augmentation, but the models scored perfectly on their own, leaving no room to measure improvement from pairing accountants with AI. So here is figure one, the first chart in this study. And the headline here on this chart, the title is Claude is faster and more accurate than accountants on medium length detail oriented accounting tasks. And we've got these black dots that represent accountants without AI. Uh they had 23 results here. And you can see that they go from zero up to it looks like maybe 90%. on accuracy and then claude alone is the purple uh diamonds 100% every single time minutes to finish the task uh claude I mean almost you know just a few minutes at most on every single task whereas the accounts took anywhere from 30 minutes to 180 minutes. Here is a chart from the study. This is figure two. The title is AI models past accountants on these tasks in 2025 and now score at or near 100%. So we've got time going from left to right from July 2024 up to July of 2026. And in the spring of 2024, that's when we got GPT40. And that was the model that first started to be able to do work to the level where we really were paying attention to it. we really started to think, oh, what can this do? And back then, GBT40 was getting on this test like 5%, maybe a few percent, it wasn't able to complete hardly any of them. And then as time goes by, we get Gemini 2.5 Pro, OpenAI 03. And you can see that basically in July of 2025, the models which you know have varying different scores of accuracy began to surpass accountants. So the average accountant score without AI assistance on this test that they do is 37%. And it's a difficult test. So it makes sense that it's it's low. And it was OpenAI's 03 that first began to beat them getting closer to a 60% GPT5 getting around 70% opus 4.8 which just was you know just uh this summer uh passes 80% and now Opus 5 100%. And all the other you know Frontier models are close to are about that. They're all over 80% now. So I mean that is a spectacular change. We're talking uh two years we went from these models being able to complete month end closed tasks like a few% of the time to now 100% of the time on this test. Now Merkur is careful to say that the results don't imply that accountants are replaceable. And there's two reasons why not. First, the tasks tested only a small set of skills that happen to be what the models are best at. Being detail- oriented, searching through files, following instructions very closely. And second, most of an account's job was absent from these tests. There's no co-workers to ask questions of. The participants came to the task cold, without context you'd normally have at a job, and that makes the task a lot harder.

There's no client communication. And there's things that the account has to do beyond what the AI is going to work on those tasks.

Yeah. And imagine, you know, you just sit down at your desk one day and you've got a new client that you have no idea uh who this is and you've got to complete some sort of month- end close task, right? That's why these are very difficult for humans to do. You don't have any history. [snorts] So you don't measure skills like client communication on this test, asking the right questions, collecting context over time. And so Mker says that these skills are going to become a larger part of accountants work as the models take on more of the detail oriented tasks. And you know what what are we talking about with detail- oriented tasks? I mean putting together um do you know doing depreciation schedules right or doing uh journal entries, right? Like taking in this data and doing the spreadsheet work that you'd normally have to do in an accounting job. Basically we're at the point where the models can do it 100% of the time for very cheap. So the work of accountants is it's just going to change so dramatically.

I think they talked about that even if the models don't get any better, where they're at today is going to change everything. Like but the models probably will still get better.

Here's an item. Here's an example of one of the items. So, there's a a a task that required uh the response to identify 4,000 of parking revenue had been miscoded as room revenue in a hotel. That's an example of this. Um what else? What's interesting too is the uh [laughter] this one's a little depressing. the the result here. This chart says Claude alone is faster and slightly more accurate than accountants working with Claude.

This is the biggest takeaway from this study. Honestly, this this is crazy. If you think about an accountant who's just working alone, right? Maybe they're it takes them 30 to 180 minutes to do these tasks. The you know, not not the best accuracy, but when Claude does it by itself, like you already showed, it's like under 10 minutes to knock it all out. 100% accuracy. But then the CPAs that are working with Claude, they're basically 15 times Claude's time. So that's pushing them into 150 minutes, which is almost as slow as the P CPA working alone, right? And then it was the uh the person working with Claude actually ruined Claude's uh percent