Frontier Models Outscore CPAs & What Clients Will Spend on Accounting

0:04David Leary: And there have been two recent studies that found that these AI assistants are actually only recommending a small group of prominent firms, meaning that most CPA firms are basically invisible to AI searches.

0:18Blake Oliver: Coming to you weekly from the OnPay Recording Studio.

0:25David Leary: Hello, and welcome back to the Accounting Podcast, your weekly roundup of news in the profession. I'm Blake Oliver. And I'm David Leary. 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 one-tenth the cost. We're also going to talk...

0:49Blake Oliver: So much in 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.

0:56David Leary: And then you've got a story about what clients are willing to pay for accounting services, 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 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.

1:17Blake Oliver: So our sponsors this week, we have Cloud Accountant Staffing, OnPay, Finance Accounting Technology Expo, and Savant Labs.

1:26David Leary: 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.

1:32Blake Oliver: 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 staffee agencies that waste your time with sales calls paperwork deposits the cloud account and 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 account and staffing you could interview someone as soon as tomorrow their boutique support ensures you're actually getting quality talent that's both available and and affordable, exactly what a growing firm needs.

2:16Blake Oliver: 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.promo.cas. That is accountingpodcast.promo.cas.

2:38David Leary: Thank you, Cloud Accounting Staffing, and welcome to our live stream viewers. we've got Tino here in the chat saying 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? Shelly happy day after deadline to those who celebrate. Great to have you with us. 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 get to see us live if you can catch a live stream. All right, let's talk about this study.

3:22David Leary: I feel like, well, I personally am not surprised by this because the recent 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 study by a startup that specializes in AI for professional services that has demonstrated just how capable these models are. That company is called Mercor, M-E-R-C-O-R, 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.

4:07David Leary: 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 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 close tasks. They can do them perfectly 100 % of the time, and they outperformed every CPA and Merkur's study at less than one-tenth the cost. What did they do?

4:45Blake Oliver: And when I say one-tenth the cost, we're talking like 21 cents a task. Like crazy, crazy cost reduction. Yes.

4:54David Leary: 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 the 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.

5:58David Leary: 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 diamonds, 100 % every single time. Minutes to finish the task, 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 GPT-40.

6:53David Leary: 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, GPT-40 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 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.

7:40David Leary: So it makes sense that it's low. And it was OpenAI's 03 that first began to beat them, getting closer to 60%. GPT-5 getting around 70%. Opus 4.8, which just was just this summer, 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 two years. We went from these models being able to complete month-end close tasks like a few percent 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.

8:33David Leary: 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 accountant'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.

9:03Blake Oliver: There's things that the account has to do beyond what the AI is going to work on those tasks.

9:08David Leary: 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 who this is. And you've got to complete some sort of month in close task. That's why these are very difficult for humans to do. You don't have any history. So you don't measure skills like client communication on this test, asking the right questions, collecting context over time. and so Merkur 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.

9:58David Leary: So the work of accountants, it's just going to change so dramatically.

10:02Blake Oliver: 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. But the models probably will still get better.

10:13David Leary: Here's an item. Here's an example of one of the items. So there's a task that required the response to identify 4 ,000 of parking revenue had been miscoded as room revenue in a hotel. That's an example of this. What's interesting too is the, this one's a little depressing. The result, here's this chart says, Claude alone is faster and slightly more accurate than accountants working with Claude.

10:45Blake Oliver: This is the biggest takeaway from this study. Honestly, this is crazy. If you think about an accountant who's just working alone, right? Maybe it takes them 30 to 180 minutes to do these tasks. 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, there are basically 15 times Claude's time. So that's pushing them in 150 minutes, which is almost as slow as the CPA working alone, right? And then it was the person working with Claude actually ruined Claude's percentage of perfect scores because the accountant would interfere and either override a number or change a number.

11:31Blake Oliver: And then the result of that task would be ruined. And that's the real takeaway from this is I think the, you think about this concept of like, let's say like AI as an assistant, right? So accountant uses AI assistant, accountant reviews use it and then there's a final output versus accountant defining rules, AI executes, automated checks happen, and only exceptions go to the CPA. The two different models are completely different now. We have accountants are slower that are using Claude side by side. I don't know how you did your taxes or I did the taxes. It's actually slower than just having the model do it.

12:11David Leary: So here's the difference in cost. Claude Opus 5 costs about$0.21. per task. Oh, per correct rubric criterion. That's what I should say. So the tasks are divided up into rubric, right? Like you have to get this right, this right, this right, this right. So Claude, 21 cents. Accountants with out AI, it was$10.35, which is almost 50 times the cost. So why did the models outperform the humans? Mercor says they believe that the tasks are fair, so you could find the right answer from the starting files. And they had built the tasks to fail models. They're full of hard-to-spot accounting conditions and difficult edge cases.

13:04David Leary: For example, in one task, a block of hotel room stays at the end of December was not billed until January 3rd after the month ended, so it was missing from the December revenue, revenue, even though it belonged there. And in another, the company background said the company does a least reclassification step every month, even though many companies only do it when preparing financial statements. Nearly every participant working without AI skipped it. They were difficult accounting requirements, but none of them were unreasonable as tests of accounting knowledge. On this question, one of the task authors said the following, The errors to be corrected in the task themselves are not rare.

13:45David Leary: Something coded to the wrong account, a bill that lands after the month ends, a commission that never made it into the books, that is a normal case. And an accountant sees them constantly. However, they continued, what is unrealistic is the setting and the scoring, meaning the model outperformance came not from the task difficulty per se, but from the interaction with the setting and the scoring. As the task author concluded, the traps are realistic, and the conditions and the scoring are what drove the results. so what diverged from reality in this test were three things the lack of context so in a real job people would have months or years of context building up to a task like this no ability to ask questions or clarify ambiguities and the lack of one-to-one expertise fit with the task so you could solve the task with foundational knowledge foundational accounting knowledge but these are testing specific concepts that may not come up regularly depending on an accountant specialization.

14:40David Leary: For example, how a nonprofit splits shared costs. If you don't work with nonprofits, you're not going to remember that from your accounting education or the CPA exam, even if you were tested on that or learned about that in school. So those are the things to note, right? Like it sounds really bad that Collada is getting 100 % on these tests and accountants are getting, you know, what was it? Very like 40%, right? But that's because accountants generally, you know, you can't know everything in accounting. So if you don't have that background or that experience, you're just going to struggle. And that is what's distorting about this test.

15:22David Leary: Seven of 12 accountants rated the tasks as realistic. So the 12 people who took the tasks, Two said they were very unrealistic. Three said somewhat unrealistic. Four said somewhat realistic. And three said very realistic. So they were kind of split on that, although more thought it was realistic than were unrealistic. But maybe it's just the ones who didn't do that well thought they were unrealistic. I mean, I don't know. I'm just speculating there.

15:50David Leary: So what else do we have in here? Yeah, I think that's pretty much it. Oh, here's a list of the task descriptions. This is interesting. We could, you know, I'm curious to know what the tasks are that they were doing. We'll just go through a few of these, right? Number one, a nonprofit runs four programs and each program's December 2024 spending was compared to its budget program by program in dollars and as a percent. Some costs like payroll taxes sat in a shared pool rather than under any one program. Those had to be spread across all four programs and the two support functions first. any line where a program ran$14 ,000 or more over budget got flagged with a note on what drove it.

16:32David Leary: So that is a task that Claude Opus 5 can do 100 % of the time. That is not an easy task. Task two was this one about hotel room revenue. The books and the booking system disagreed about the December room revenue, which also threw off commissions owed to its online travel agencies for the month. And both gaps had to be tracked down item by item with the amount in the account or platform for each, then fixed with correcting journal entries. The net effect of both fixes on December, then the net effect of both fixes on December income. Task three, a venue's events in December brought in more revenue than budgeted, but the average per event came in under plan.

17:20David Leary: Actual and budget with the gap between them were needed on the number of events, the average revenue per event, food and beverage minimums, and the mix of event types by count. Last came an explanation of the difference backed by revenue and event count for each type of event. Again, these are not easy tasks to do. There's one here about three companies in one group, each renting a property. There's a holding company with an office, a trucking division in a terminal yard. and a logistics division has a warehouse. And under lease accounting rules, each lease sits on the books two ways. There's a lease liability for what's still owed and a right of use asset for the value of using the space with both coming down over time.

18:05David Leary: And for December, each lease schedule was rolled forward a month, working out how much of the asset got used up and the drop in the liability. The numbers went into the December entries for each company's QuickBooks file. So again, these are tasks. that all 12 CPAs were given. They all had accounting-specific degrees. Eight out of the 12 had a master's degree. They all had U.S. CPA licenses, and their average experience was 5.4 years. And six were in public practice and six were in industry. Half of them had worked at a big four firm. So quite a conclusion here. I don't know, quite an impressive result from Claude Opus 5.

18:53Blake Oliver: I mean, I think the one observation conclusion is this whole notion that a CPA that's also using Claude is actually not very efficient. Like, that's an issue. And this goes back to, I know we've talked about that. Hey, you have to document the process. So, I just want to challenge you on that a little bit, right?

19:09David Leary: So it was like in a small set of like 20-something tasks, a few times, right? Like four out of them, four times, four or five times, the CPAs actually made the model worse. They messed it up.

19:28Blake Oliver: Yes, yes. But in general, everything they did took 15 times as long as Claude.

19:32David Leary: Well, and I think that's probably because when you use Claude in the manual mode, you have to read everything it says and then approve it. And then read everything it says and approve it. And so that takes a long time. That's why it is. They're slow. It's going at human speed as fast as we can read and analyze.

19:54Blake Oliver: Yeah, because it's like a human oversight problem. The humans are trying to oversight everything. but so if you kind of step back and think about this like you probably need to create brand new ai native workflows at your firm because if you just give your staff a hey i want to use ai as you're working you're probably not going to see any roi in the you're going to be like we got all this ai i gave it to my staff they're not any more efficient you really have to completely change your workflows to an ai native first workflow ultimately so what is the ai good at the ai is

20:27David Leary: good at taking all this information and accomplishing the task when there's like clear explanations of what to do there's a there's you know exactly the AI knows exactly what it should do and it has all the inputs it needs to do them the problem is a lot of times the AI doesn't have all the context so that's why these tests are not fair and what I'd be interested to know is is well I guess the next question then is how do you actually like take advantage of this capability like using Cloud Opus 5 to do the work in real life, not just in a test scenario. And I feel like the way to do it is to use model orchestration or agent orchestration, where first, right, you have to gather all of the information you need to complete the task.

21:15David Leary: And actually, even before that, you have to define what the task is. So there needs to be a model or not a model, an agent guided by a human that like defines the task. Yep. And then there needs to be an agent that gathers all the information necessary. And the human is going to have to be in that loop to help. And then you need the agent that actually completes the task. So if you combine the human who's going out and talking to people and getting the information and working with the other humans that have this information and bridging systems and gaps, and then you've got that human CPA working with agents that can help to define the problem and the solution and then do the work.

22:03David Leary: That's how you get to automated accounting. right and I feel like if we've gone this far in a couple of years we are going to get there in the next few there are companies that are going to figure out how to build that orchestration layer around these these models that can now do the work once they have all the information

22:29Blake Oliver: yeah defining the rules is the key right because if you do it the other way you're just going to be questioning the results you got. Hey, are you sure that number's right? And you do that dance back and forth.

22:40David Leary: Dustin Pruitt said in our YouTube live stream, I think David finally hit the nail on the head. You not only have to use AI, but build it into your process so it actually saves people time. And yeah, David, that's so true that like currently the way we're using it now, it doesn't actually save us that much time. It still takes a long time to do these tasks the way we're set up. And I find that too myself. It's like, I am doing so much now with Claude, but I find myself sitting at my computer having to like approve and wait and approve and wait and answer a question and wait. And so some days it's really not any faster and figuring out how, when I can like, let it go.

23:29David Leary: And when I can, when I have to watch it is the hard part, Because at this point, I'm not sure when I can let it go and when I need to watch it. And the agents are not so great at knowing when to continue and when to stop. That's the judgment. How do they know when they need to stop and ask a question versus when to just continue on and not bug me? That is something I haven't figured out how to solve it.

23:54Blake Oliver: That's the bridge that gap. When does something become an exception that you need to be involved in versus just getting in the... Because really, every time it waits for you or wants feedback along the way, that's just going to slow it down. The more you can take you, the human, out of the human involved with AI, and you only get involved when there's an exception, that's where you're going to get those efficiencies.

24:17David Leary: David, I am going to thank our next two sponsors while you get ready to share this story about what clients are willing to spend on accounting services. thanks to on pay for sponsoring this episode of the accounting podcast are you tired of payroll headaches getting in the way of the client experience you want to deliver manual workflows creating bottlenecks compliance nightmares and endless support calls that go nowhere there's a better way for your team and your clients on pay is the payroll partner that accountants and bookkeepers actually love why because it's easy to use packed with value and backed by support that actually supports you.

24:54David Leary: Their team gets rave reviews for being fast, expert, and actually reachable when you need them. OnPay handles the heavy lifting. You get a dedicated onboarding coordinator who sets up worker profiles and transfers year-to-date data from previous providers, all at no extra cost. Their seamless QuickBooks and Xero integrations eliminate manual journal entries. And they support any type of business you serve, farms, restaurants, nonprofits, you name it. OnPay can handle unique requirements without adding complexity. OnPay keeps pricing simple too. Everything your clients expect from multi-state filing to off-cycle pay runs is included.

25:29David Leary: There's no hidden fees and no surprises. To book a demo, head over to accountingpodcast.promo.onpay. That's accountingpodcast.promo.onpay. Also, thank you to Fate for sponsoring this episode of the podcast. If you're a regular listener of the Accounting Podcast, you're always looking for practical ideas to build a smarter, more efficient finance function. That's exactly what the Finance and Accounting Technology Expo was built for. On November 8th and 19th at the Javits Center in New York City, finance leaders from across the country will come together to explore the latest accounting and finance technology, compare solutions side by side, watch practical product demos, and learn from peers who are navigating many of the same technology decisions you are.

26:17David Leary: Whether you're evaluating AI, automation, ERP, FP &A, AP, or closed management solutions, FATE is designed to help you get a comprehensive market scan in just two days without weeks of individual vendor meetings. Use code ACCOUNTING50 for 50 % off your registration. That's code ACCOUNTING50. All capitals. Head over to accountingpodcast.promo.fate. That's accountingpodcast.promo forward slash fate and use code accounting50 for 50 % off.

26:53Blake Oliver: So have you heard of Sam's List?

26:56David Leary: Is it like Craigslist?

26:58Blake Oliver: In a way, but it's just to let small business owners find accountants. It's like a search engine type thing. Well, they analyzed, they have a quiz when people go to the site, And they analyzed the answers from 1 ,926 people in the US between January of 2025 and October 26 that were trying to find an accountant, a bookkeeper, a fractional CFO, or a financial advisor. And they just have a lot of learning. So some of them, it's really interesting. More than half the people, they want help within the next two weeks. So if you have a client contact you, right, if you have a, you know, new onboarding process, people want to make a decision and use your firm in two weeks.

27:40Blake Oliver: So it's really important that you do that. And 78 % are with a month. Half the people hitting that site are hiring a financial professional for the very first time. And 36 % are just replacing somebody they already have. And the main reason why is they just don't feel like the existing one is meeting their needs at all. Now, the budget, this is what's interesting. So people will budget about five grand for accountant as long as possible. But then once they hit a million dollars in revenue, they're willing to spend more and they'll spend 15 ,000. So it really jumps, but it's all based on that that initial revenue size.

28:21Blake Oliver: They will stay, you know, 700 ,000, 800 ,000, 900 ,000. They just do not want to spend more than five grand until they tip over that.

28:28David Leary: Until they hit a million and then the budget triples. Yeah.

28:32Blake Oliver: Then it triples. Um, so that's interesting, David,

28:35David Leary: because I've always felt that way. Just my gut feeling after working in accounting with tiny little clients all the way up to mid-sized ones as I change firms and my business change, my feeling was working with clients doing under a million dollars a year in revenue is so hard because they just don't want to spend. But then somehow once they get over that marker, that that one million dollar mark then they are willing to then and it's substantially more and so this backs that up so basically i tell people like whenever accountants were asking me like who should i focus on with my firm i say unless you have like a unless you have a really really simple firm that just is the same thing for every type of client where you can really eke out economies who scale and get good margins, don't work with clients under a million dollars.

29:37Blake Oliver: I think that's always easy to say, right? But there's only a finite number of those clients out there.

29:44David Leary: Well, no, but that's the thing is there's so many clients under a million. Like the vast majority of small businesses in this country are under a million. And, you know, solopreneurs are in there. But even the ones with employees, there's so many businesses that have just like a handful of employees and an owner or a family business or something like that. you know, like fewer than 10 people that are at or below a million, there's so many of them. And so they're looking for help all the time too, but they're cheap and they don't want to pay. And this backs that up. And so like the only way to actually build a business that is like saleable someday, sellable, saleable is if you only focus on like one type of client.

30:25David Leary: So it's like, you know, just real estate agents. If you only did real estate agents, then yes, Yes, you could build a business that could scale. But other than that, no, don't do it.

30:37Blake Oliver: So bookkeepers are the most urgent hires. So that's do they want immediately when they're trying to hire. They need bookkeepers right away. But then it talks about also kind of what they want help with. And this is interesting. 74 % are looking for tax prep and filing. 69 % tax strategy and planning. 28 % bookkeeping. 15 % sales tax compliance. 13 % payroll. role, 9 % R &D tax credits, 8 % IRS audit help, 6 % international tax. Now, I find it interesting that so many firms lean into this. Well, we do international, we do R &D credits, but people aren't shopping for that. So many firms, they lead with that and that's our specialty, we do that.

31:21Blake Oliver: But most people just want tax prep and filing, tax strategy and bookkeeping. I mean, that's the three services that the core of your services should be. Yes, you could specialize in the other stuff, but the demand's just not there. People aren't looking for that.

31:35David Leary: Here's something in the survey that stands out to me as well. It's the percentage of these small businesses that are willing to spend more than$20 ,000 on accounting and tax. It's very small. and 20 ,000 is important because I don't know if I'm providing like accounting services on a monthly basis and a tax return I think it's really hard to do that well it's really hard to do that well if you um what am I trying to say try to do that well for that budget right because$20 ,000 a month is less than, you know,$2 ,000,$20 ,000 a year is less than$2 ,000 a month.

32:22Blake Oliver: Yeah. So, and, and I don't know, like I tried doing like monthly accounting services for$1 ,000

32:31David Leary: a month. And if you're going to do it right, it's really challenging. I mean, unless it's like completely automated sort of work, you know, I mean, I started out doing it for like$500 a month. And as my business grew, like$1 ,000 a month became the minimum. And I mean, that was years ago now, right? So now it's like, how could you do it like profitably for less than$1 ,500,$2 ,000? I don't know.

32:59Blake Oliver: Only if the AI is doing all the work. And you're doing it for$10. But right now it's not really speeding us up, right? Yeah.

33:05David Leary: Yeah. It's not automated. So anyway, that's a long way of getting to this stat, which is that for businesses between$500 ,000 and a million in revenue. So half a million to a million in revenue, which is a pretty decent-sized piece of the market. Only 13 % of them are willing to spend$20 ,000 or more. So what does that say? It says that you should be turning away, if you want to make$20 ,000 or more, and you're getting 10 people coming to you in that range asking for your services, you are going to be too high of a price for like nine out of 10 of them. So your close rate needs to be like 10%.

33:49David Leary: If it's more than that, you're probably, probably, I mean, maybe it could be more than that, but like, you should just expect that nine out of 10 businesses you talk to in that range are going to be not willing to pay$20 ,000

34:05Blake Oliver: or more not not not even that they're not willing to pay uh they plan so there's they have a stat 77 of early stage startups these are startups but they plan to only spend 5 000 or less on their accounting services like in their business model that's kind of what they budget so so it's not even like an afford thing it's it's a mental uh it's a mindset right people like smaller companies are just like i it doesn't make sense to spend more than five grand i don't know where this mindset came from, but I kind of understand where it's at.

34:39David Leary: All right, David, let's go ahead and thank our next sponsor. And that is Savant Labs. If your finance team is using AI for drafts and research, but still closing the books by hand, you're in the same spot as most enterprise accounting teams. Claude and Copilot handle the easy stuff. But what about the close, the reconciliations, and the tax provisions? Still manual. Still spreadsheets. Still eating up days every cycle. The real issue isn't effort. It's trust. General purpose AI gives you a different answer every time, and an auditor can't follow the trail. That's why Savant exists. You describe a finance task in plain language, and Savant turns it into a governed agent.

35:22David Leary: It pulls, cleans, and reconciles your data across CRPs, spreadsheets, and PDFs with over 500 connectors. Every workflow runs the same way every time. Same input, same output, every cycle. Data lineage, approvals, and SOX reports are built in automatically. The result is up to 70 % less manual work and 80 % fewer errors. And here's the best part. Bring Savant a complex use case that's eating 20 hours per month or more, and they'll build you a custom agent for free with a 30-day trial. To learn more about Savant and start building your first agent, head over to accountingpodcast.promo.com. savant that's accountingpodcast.promo forward slash savant and please when you go learn about our sponsors use those links that helps them know that you heard about them on the show gives them something they can track and gives us something we can track and our sponsors really appreciate it appreciate it and it helps us grow sponsorships on the show and that's how we make our living.

36:25David Leary: So thank you.

36:26Blake Oliver: And if you want to see a demo, Savant, we're doing an expo with them next week. So just go to the Earmark app and you'll see the Earmark expo and register.

36:35David Leary: Going back to the live stream chat, Tate says, Blake, you look like an AI with that background. Y 'all should try to write a script and do an episode 100 % by AI. Normally I don't use a virtual background, but today I am. And that's because I've moved into a new home. and I'm in my new podcast studio. It's amazing. It's a separate structure from the main house. I have like, it's its own thing. And I'm going to show you what it looks like right now. I've got a extra mattress there propped up against the kitchenette. It's great. It's got a kitchenette. It's got its own bathroom. I can just like hide out here and podcast.

37:16David Leary: Yeah, a table full of, oh, there's a microphone. My original Yeti microphone. from back in the day yeah so that's what's going on that's why i'm all disorganized today if the show seems more discombobulated than usual that's the that's the reason we've got boring accountant here with four coffee emojis great to see you boring account i unfortunately still haven't found all the parts to my coffee maker so i'm having to go out and get starbucks every day and i i should just get one of those like giant like what do they call them the takeaway boxes of coffee yeah yeah because the the one the one latte is not doing it for me thank you tate thanks for the congratulations tate also said think of our current ai setup as cameras on phones in 20 in 2007.

38:06David Leary: cameras on phones in 2007. theoretically they are but not good in about 20 years they're going to get really good like phone cameras have yeah except i agree completely except for the timeline. I think it's going to get really, really good in like five years. I mean, just look where we've come. It's been what, uh, I can't even, I can't remember when I started playing around with chat GPT, but in two years we've gone from these models being basically hopeless at monthly enclosed tasks and only able to do tiny little things. Like, I mean, we were just drafting emails. Right. And now I got to tell you, David, about something I did, or I'll tell the listeners too.

38:47David Leary: so uh this is this is just how powerful opus 5.5 is um i'm using on like the medium setting because you can do like a i don't know like different levels of thinking so we uh we have this like ai course drafting workflow that i built years and years ago and i did it uh you know like we're talking like two years ago. And the way I had to build it two years ago is literally I figured out how to draft AI courses from transcripts. So like the transcript of this episode, I would put that into ChatGPT and then do a series of prompts to get it to come up with the description for the course, the glossary, the index.

39:35Blake Oliver: And you had to basically force it into breaking things up in chunks because it couldn't even handle a full transcript, right?

39:41David Leary: I had to wait a while for it to be able to handle the full transcript. I think before that I had to chunk it. Eventually the context window got big enough to hold an hour long transcript. So that was amazing. So I'm talking about that point. The context window gets just big enough. And what I had to do was basically upload the transcript and then do a prompt for every single piece of course data. So every metadata item, the index, the glossary, each review question I had to do as a prompt in a new chat, right? I had to start over because the chat context would fill up. You couldn't do this all in just one go.

40:21David Leary: And so it was like dozens and dozens of prompts. And I just saved those prompts and iterated on them and improved them. I had just this doc of prompts. And so when I made a course, I would just be prompting and prompting and prompting so many of them and then I built that into our workflow software to automate that prompting and that has been how we've drafted courses for all these years like two years now just the same method and now with Opus it's changed you can now do one-shot prompts they give it everything almost everything in a single prompt and then opus delegates the work to sub-agents so in claude now the co-work feature of being able to create sub-agents it just orchestrates them automatically so it'll it'll then delegate the work of creating the index to a sub-agent and the glossary to a sub-agent in each of the review questions and quiz questions and then it will review the output and And then make sure that that matches what it was.

41:30David Leary: It creates like a rubric and grades it. And if it doesn't get it right, it does it again to a different subagent. And so I wanted to do that. But it took me months to build the first workflow. And I just haven't had time to redo it. So I just asked Claude to look at the workflow in Zapier and to rebuild it, given all the new capabilities. And basically over the course of like a day or two with it going back and forth with me, it recreated the whole thing as an entirely new workflow with optimized AI prompts using the best models. And it plugged it into our software. This is something that took me months to build.

42:19David Leary: And it rebuilt it. And it's amazing. The quality is insanely high. so i don't know why i started telling that story i think it's because

42:30Blake Oliver: things are changing so fast things are changing right yeah and remember what we just talked about muse facebook or meta's ai agent i would say it's it's ai's microwave moment or whatever it was what's that three weeks ago four weeks ago right well um dual entry has their own accounting ai benchmark. And Muse's newest model, Wonder, got number two among 57 AI models tested on their AI accounting benchmark. So we're going from a brand new model, four weeks later, it's doing accounting work. Now, it's struggled a little bit more with operational tax, like accounts receivable and transaction classification, but it was nailing reporting 97 % and reconciliation at 92%.

43:15Blake Oliver: 22%. So, but you just have how fast these things are changing. That model just launched. Nobody even used it before four weeks ago. And now it's already doing 90%, 97 % reconciliation stuff. Do you want to talk about, I get a couple of choices here. Ramp spent 1.5 million a week on tokens. You want to talk about that?

43:41David Leary: It was in one week, they spent 1.5 million. Yeah.

43:44Blake Oliver: So in May of 2026, Ramp had a one week situation in May that they spent 1.5 million on tokens as a company overall everywhere. And everybody freaked out, right? It's just too much money being spent too fast. We have no control over this. And a lot of those problems are is because people use the most expensive, best model to solve an easy task, right? So you're overpaying to get that task solved. So what, and the other problem was you don't have a lot of context or control. They can't see where the spend's happening. Everybody's using different models. Even for us here, like I do some stuff in Claude, I have something else in chat GPT, I'm using Muse, I'm using Grok, I got stuff everywhere.

44:27Blake Oliver: So what they built, Ramp launched a new product that's called Glass. And what it is, if you think about it, that's the AI model inside of RAMP. And everybody in the company uses this one AI model called Glass, even though it's not even an AI model. It's more an orchestration model, right? It's like an agent tool. It's an agent, and then it uses all the other – it's a router to –

44:56David Leary: Okay, so it's not a – it can use whatever model, and it routes the prompts. It routes it, right?

45:01Blake Oliver: And so what they get out of this is they have control on how tasks are being executed. They have insight to spend that's happening on tokens. And then B, they can control what task goes to the right spend, right, or the right model to be used. And then the bigger thing of this, which I find really interesting, is now everybody's stuff is in one place. So now all the models and everybody has context. So if you did work a month ago and I'm doing something that's adjacent to the work you did, the work I'm trying to do is going to use some of the work you did before. And that's what's really interesting about this.

45:40Blake Oliver: It's like orchestration for the whole company's use of AI models.

45:45David Leary: So they can learn how people are using it and then they can gather insights from that and decide how to improve.

45:52Blake Oliver: Well, no. Let's say you worked on a quote to somebody. Maybe you did a proposal. And now I'm going to do a similar proposal. It just knows about your proposal.

46:05David Leary: It has access to all the chats.

46:07Blake Oliver: All the chats. It's everything in the whole company. It's almost like a super internet for the company. All the company's information now is in the same product. But they're making this available. Well, there's an early invite list you could put an email address on if you go to ramp.com slash glass. but it looks like this could be a product that would make a lot of sense for accounting firms because ultimately unless you're a huge company and you have somebody orchestrating all this for you most accounting of smart accounting firms don't have that you want to just buy an off-the-shelf product tell all your employees to use the same product and not manage your tokens etc and we get in this situation here right I did some work on product design and Claude but I had a conversation with you in our email program, and I had to go to that email program's AI to get the results, because I forgot what you and I talked about, and then go take that to put it back in my conversation in Claude that I'm designing the product with.

47:04Blake Oliver: It's just a lot of dancing.

47:06David Leary: Oh, well, by the way, David, you can now connect Missive to Claude. So you can just have it go grab all the contacts you need.

47:12Blake Oliver: Yeah, but now you have it all. This is going to be an off-the-shelf rollout. It looks like, and I think it's really smart, but it's funny. It's all driven because they were spending too much money on tokens.

47:27David Leary: Let's talk about how AI is changing search. So accounting firms have been spending big on search engine optimization for decades now. and it's crazy how much blogging was done over decades to get people to find your website on google and it worked for me it was great i did a whole series of blog posts on zero versus quickbooks online when i started my firm nobody else had any content out there like that because zero was new in the u.s and i basically got everybody searching for zero finding me in the top 10 results, it was the marketing engine for my firm. So it works if you do it right. But AI is changing that because people are no longer going to Google into the search box and searching for something and going through the pages of results.

48:20David Leary: They're just asking ChatGPT or Claude, or even when they go to Google now, they're getting Gemini answers. So that is how they're finding accountants. They're asking AI, who in my area can do my taxes or who should I hire to do my taxes or my books or that sort of thing? And there have been a variety of people studying, how is this working? How is this changing marketing for accounting firms? And there have been two recent studies that found that these AI assistants are actually only recommending a small group of prominent firms, meaning that most CPA firms are basically invisible. to AI searches.

49:01David Leary: There are some smaller practices that outperform through stronger digital positioning. So this first study was done by a marketing advisory firm called Expert Press, and they asked 100 buyer questions about tax, audit, advisory, risk services across ChatGPT, Claude, Gemini, Google AI, Perplexity, Grok, DeepSeek. Each question was tested five times, so they got thousands of responses. Just 25 firms generated 87 % of AI search results. So 25 firms, basically all large firms, or mostly all large firms, got AI search results. Who are they? A lot of these names you'll know. RSM, BDO, Baker Tilly, CLA, Grant Thornton, Forvis, Mazers, Kohn Resnick, Eisner Amper, Sherry Becker, Crow, Plant Moran, Witham, Seabiz, Armanino, Wipfleet, Aprio, Citrin Cooperman, Ide Bailey, PKF O 'Connor Davies, BPM, Barry Dunn, Sickich, LMBC, no, LBMC, the Bonadio Group, and Doran Mayhew.

50:09Those firms got 87 % of the results.

50:13David Leary: If that was a Google search thing, that would be incredible for them to all that few firms to be capturing all that. Top 10 firms captured 61 % of the mentions, right? RSM was first getting mentioned almost 30 % of the time in these questions. And like the bottom 80 % of the top 100 firms were mentioned only 3 % of the time. Eight received no mentions. Interestingly, why, why are these firms, right? Like why, why are some firms getting mentioned more than others?

50:46Blake Oliver: Because I just did a search right now on Microsoft Copilot. I was like, hey, I need to get my tax return done first. S-Corp and my personal return. Who should I contact? And it basically gave me Tucson Tax Team, 1-800-Accountant, and TurboTax. So, yes, why is that the result?

51:06David Leary: Well, so revenue correlated strongly with visibility in this survey. So the median firm in the highest revenue tier appeared about 19 times more often than the median third tier firm and over 350 times more often than the median firm in the lowest tier. So the bigger you are in terms of revenue, the more likely you are to show up.

51:26Blake Oliver: So somewhere AI has it in its head that because it went and read the list of top accounting firms and knows the revenue and said, these must be the best.

51:35David Leary: I wonder if that's because like in accounting, we rank firms by revenue. like accounting today's top firms it's all revenue based maybe that's because we've done it to ourselves um but it's hard to say because large firms naturally produce more online evidence of authority they make more content um when you account for that right how much content firms are producing small firms can perform well you have to have broad authoritative and specific digital footprints that give AI systems evidence connecting a firm to a buyer's question. So what would that be? You need to have, obviously, a website with pages that explicitly name the geographic market, describe the buyer-specific work in depth, have been updated recently.

52:27David Leary: Those are strategies that also used to work for the SEO era. Other useful evidence, government and client documents, industry associations, commercial directories and ratings databases. The models generally pass over Reddit, Quora, and Wikipedia and rarely cite news unless there's a major list or ranking.

52:50Blake Oliver: So is there like a learning? Like how do I get my firm to be recommended?

52:55David Leary: You have to build clear authority around a niche, location, and problem set across your own website and third-party sources. So that's actually not very different than building SEO.

53:06Blake Oliver: Yeah, the niche makes so much sense. I'm a bicycle shop. Boom. It's going to already have a filtered list to recommend.

53:13David Leary: I think what will not work as well in the future is the location-based SEO. So getting Google Maps listings for multiple offices in different cities in your region, even if you don't have an office there, getting like a fake one or a post office box, Like that's a common strategy because search tends to give preference to geography. Google in particular, people search on Google. It's going to want to recommend a firm that's near them versus one that's far away. But AI doesn't seem to care about that. So AI is going to care more about whether or not you fit with the profile that the AI has built around its user, the memory that it has created.

53:57David Leary: So the particular type of business. and that's always been a really good content marketing strategy because it was good once you got clients on your website you wanted them to see that they that you specialize in what they do that you specialize in clients like them and that was always the reason to do good content marketing it wasn't necessarily top of funnel it wasn't to get leads in the door it was to build authority so that then they perceive you as being an expert that they want to contact versus a generalist firm that doesn't specialize in what they do. But now that strategy is actually becoming even better because it's becoming a top of funnel strategy because the AI is going to your website and doing the thing the human used to do.

54:43David Leary: The AI is going to the website and looking at what you specialize in in determining whether or not you're a fit for its user. And so then that becomes top of funnel. So basically, content marketing, podcasting, making videos on YouTube, blogging, social media posting on your particular specialty in your niche, demonstrating your expertise is even more important now. Which is kind of how Google always judged everybody to begin with.

55:16Blake Oliver: It's just now it's just happening faster, I guess. Even more.

55:21David Leary: anyway David I have to run that's all the time we have I have to actually literally sign off get out the door go to an appointment so I want to thank everyone who joined us live today great to see you sorry if we didn't get to your comments don't forget you can earn free CPE for listening to podcasts like this one with the earmark app go to earmark.app in your web browser or get the free earmark app on the app store create your free account earn one CPE for free per week subscribe for premium content and to take unlimited courses and to support the work that we do. We really appreciate our subscribers.

55:57David Leary: Get that subscription going as you get into CPE season now that the tax deadline has passed. And we'll see you around here next week. Thanks, David. Bye, everyone.

56:13Blake Oliver: Time for the classifieds.

56:19Blake Oliver: Can you defend your client's Shopify margin number? ProfitRoot turns a Shopify orders export into a reconciled margin report. Every figure traces back to the source file, every coverage gap is named, and nothing is estimated to fill a hole. There's no sign up to try it. Walk the preloaded sample store or upload an export and the first report is free. Built for accountants and fractional CFOs. Try it at tryprofitroot.com.

56:47Blake Oliver: Want to get the word out about your newsletter, webinar, party, Facebook group, podcast, ebook, job posting, or that fancy Excel macro you just created? Why not let the listeners of the Cloud Accounting Podcast know by running a classified ad? Hit the show notes for the link to get more info.