Transcription
This is the best AI agent for accounting firms that I've ever seen, and frankly, it isn't even close. It's now available on Mac and Windows. It's called Claude Co-work, and we just benchmarked it on a ton of accounting firm work, from bookkeeping to tax to automating Excel and QuickBooks. Like, real work that you do in an accounting firm, and it is stunningly good. If I'd had this in my 40-person accounting firm, I would have changed how everybody in my firm worked.
The problem is, I can't yet trust it with sensitive client info yet. I can't yet. And we're going to talk about that. But you need to see just how good it is today because it's not an exaggeration to say that it's already capable of automating a lot of the work that we do. So today, I just want to show you where AI agents are at. I'll show you a bookkeeping example, a tax prep example. We'll talk about the cost to use it, the security behind where it's at right now, and how it could even automate your QuickBooks.
We are at a really wild moment right now. When ChatGPT first came out a few years ago, I couldn't sleep because I'm a nerd, but also because I almost couldn't find the boundaries of like, what can't this AI tool do? That was 3 years ago. Right now, this is the first time since then that I honestly haven't been able to sleep. And this isn't like slimy YouTuber hyperbole. It is wild.
Let me just show you first with a bookkeeping example. Let's say you got a simple client and you're like, I really don't want yet another QuickBooks subscription for a very simple accounting file, right? Maybe it's an itty-bitty rental, or you do tax work. You get a lot of these these itty-bitty like compliance books where the client just gives you some stuff, you throw something together. We used to use QuickBooks Desktop for this. I think all that's about to change.
So, here's a spreadsheet that I use for simple bookkeeping. You just put the transactions here and you classify it with a letter here. I know you're watching this on your phone in the toilet. Just trust me. These are all the categories up here, generally mapped to the Schedule C. You can do a rental version of this as well. But take a look at this file. I've got three months of bank statements, 67 transactions in total. I'm going to put this in a cloud. Enter all the transactions from these bank statements into my spreadsheet template and classify them to the best of your ability. All right, that's going to start working now.
Quick tour of Co-work. This is a local app that you can install. Originally came out on Mac. It's now available on Windows as well. You see on the top here, we've got the Co-work option selected. On the left, you can select chat. That's just like a chat assistant like ChatGPT. On the right is code for like developers writing code. Co-work, it is the safe spot for normies like you and me. Up here, it builds this little like to-do list. You can see the context, or basically what it knows, what it looks into. You'll see skills referenced here. We're going to circle back to that. Here's the output, just like that. 67 transactions. If you look at the detail here, how it describes how it classified things, that's pretty intelligent. And at the bottom of the spreadsheet here, everything ties out to the final balance on that third month's bank statement. Pretty cool.
But can it do 670 transactions? Here are 12 months of statements, buddy. I'm going to drop these bad boys in here and give it the same prompt. And off it goes. While that's running, again, don't use this with client data yet. We're going to talk more about that in a bit, but we almost didn't make this video because we're like, we don't want people running off with this. But also, I think it's important to see how like, where this is at right now, even though we can't use it with client-sensitive stuff yet.
But how we got here is two main things. First, some phenomenal new AI models from Claude, namely Opus 4.5 and this Opus 4.6 that we're using, models that are capable of more than any other models that we've had before. But second, Claude Co-work as a novel, uh, what people would call harness, basically, because we're not simply talking with a large language model. When I ask it to do something, it's spinning up like sub-agents and kind of orchestrating all these little jobs to get done. And Co-work has figured out how to do this in a way that's like useful, productive. And the combination of that with really good new AI models means like, this sucker is, it's really good. So, I'll show you what this looks like at the end. 670 transactions. It took 12 minutes, but the output, again, it's very good. Down to the bottom of the spreadsheet. The ending balance matches the December ending balance in the bank statement. The classifications, as far as I can figure, pretty much what like a human would do. And again, remember what I said. You shouldn't be putting your client's transactions into this for reasons that we're going to get into.
Then why are my clients' transactions in here? These bank statements generated with AI. Yeah, they look convincing, right? We've even got a few different styles here, like a big bank style, little baby local bank style. Honestly, it's really easy to do. But I built this whole transaction generation and bank statement generation thing into a skill. And skills are a huge thing in Co-work.
So, to generate the bank transactions, I told it, I'm an educator that teaches bookkeepers how to do small business bookkeeping, blah blah blah. How can we make some transactions that look real? We did some back and forth and it's like, well, you're going to want multi-month continuity. You want weird stuff like NSF fees and return checks. Make the statement descriptions not be very helpful because they rarely are. It'll pick the business's industry. It'll factor in seasonal cycles for the business. It'll give the business an actual real-world location and use local to that location vendors for stuff that you would expect to be bought in person. Like if you're doing a doughnut run, it will go find a local doughnut shop. All of that in the transaction generation. It is very good. And then I simply say, make it into a skill. And now in any conversation, I can just hop in and say, "Create nine months of transactions for a taxidermist." 450 transactions in total. That kicks off down here. You'll see it then automatically uses the transaction creator skill. That's that skill that we made. And when it's done, bing, it'll give me a CSV of really convincing transactions. Then I made a second skill to put those transactions onto bank statements. And voila, pretty darn realistic bank statements.
So skills, they are very cool, a big part of Co-work. And I'm going to give you every skill from this video if you want to play with them. You wouldn't believe just how powerful they are. Case in point, we just ran 670 transactions through this, but why stop at 670? Let's do 6,700 transactions. Trent, Trent, you got to see this. So those 670 transactions took 12 minutes. So 6,700, like, ain't nobody got time for that. I don't think he can say that.
So after Claude finished those transactions, before I said, "Why did this take you so long?" Let's create a skill to help you complete this Excel ledger template faster. And it turned out that the classifying of every transaction was the slowest part. It would go out and it would think about every transaction, maybe do a bit of research. What is this vendor? That took forever. So, we made a skill that creates a vendor map on the fly as it's classifying transactions. So, it doesn't have to think about Home Depot every time. It thinks the first time it sees a new vendor, but then every subsequent time it autoclassifies it to what it picked previously. So, I packaged this into a skill called "Ledger Template Completer," then came back and reran the 67 transaction test and said, "Use the Ledger Template Completer skill." Instead of 3 minutes 36, it took 3 minutes 5, which was disappointing. But the 670 transaction test went from 12 minutes 10 to 3 minutes 12, people. So we ran the same thing across 6,700 transactions. Here's the output. It, uh, it worked. I don't know what to tell you. Look at this. It took 7 minutes and 29 seconds. And again, I know you're watching this on your phone in the turret, but trust me, it processed all these transactions, 6,700 of them. If I go to the end, the balance ties out to the last bank statement, ending balance right here. By the way, look at this monster of a bank statement. Look at all these pages. This was all created by AI. Not just this statement, 12 months of these statements. And then I gave all these statements to the AI, said, "Code them in this spreadsheet." It did all that in 7 minutes. And the classifications, like if you look at them, and if you look at the logic that it explains here, here are the things that you may still want to look into. Here's how I classified stuff. It's really well done.
Transactions were in that thing you did last week. Like 2,500. 6,700 transactions in 7 minutes. That blows my mind. But there is more, gang. This video, we are just getting started. Receipts. That's 41 of them. It's a good joke. What is? I hope he makes it some more times. So, I had it link all 41 receipts to its associated transaction in that Excel file, then renamed the receipts, uh, in this file name format. It ran for a minute and 10 seconds and gave me this linked Excel file back. The linked ones are kind of blue here. I changed the formatting, and when I try to click through to the PDF, Excel gives me like eight pop-ups. Are you sure you want to open this file? Yes, I do. Thank you. But it linked 41 of those transactions. It renamed. Look at this. It renamed all the files as I asked it to. That took like 60 seconds, gang. And these receipts. Look at these. Notice anything? These receipts. I generated these with a receipt creator skill. Again, you're going to be able to swipe all these skills. But look at these beauties. But by the way, these receipts, they're actually images in the PDF. It's not text in the PDF. So Claude had to go out and OCR every single one of these. And it, it did all that in the minute that it took to process all this.
Now, I wanted to throw more receipts at this, right? So, for example, what about 410 receipts at the 670 transaction spreadsheet? Nice. The trouble became that generating 410 receipts with AI was taking so bloody long, I just wanted to move on with my life. But it kind of makes you think. Receipt management apps. Are those kind of like going out of style? Will AI just be able to handle a lot of this for us? A bit later, we're actually going to try the same thing in QuickBooks Online. Can we get this tool to go out to QuickBooks Online and actually attach the receipts to the transactions inside of QuickBooks? But yeah, receipt management apps, be interesting to see. Actually, if you want all my app recommendations for accounting firms, they're linked right down below in the video description. Used by over 8,000 accounting firms now, the only apps I recommend firms use.
But let me next show you a quick tax example. Even for non-tax folks, you're going to find this fascinating. Now, we're going to cover how much all this you're seeing costs. We'll go deeper on security, but a quick tax example. I've got three client tax files here. This one's a simple one. Just three docs, a W2, 1099 interest, and some student loan interest. Second tax file is a little beefier. 15 tax docs, taxpayer, spouse, normal stuff, a big brokerage statement from Fidelity, 22 pages in total. And third, a big boy, 30 tax docs, all sorts of stuff, multiple brokerage statements, HSA distributions, uh, 401k rollover, all sorts of stuff. 57 pages in total.
Now, I've got this 1040 workbook right here. It's like a giant workpaper to put all the tax info in to tie to the tax software. Make sure that I'm not missing anything. By the way, this is Excel 1040. It is a glorious workbook. Basically 60 tabs of input tabs and then tabs recreating 1040 schedules. Grab it at excel1040.com. It's totally free. Those three tax files. I wanted it to enter the tax file into that Excel workbook. I'll just cut to the chase because Co-work did all three without even breaking a sweat. The biggest one, the 57-page one, took almost 30 minutes. And I, I did some work on building a skill for this because I think it can be made faster. But that was on its first try. I did it in about 30 minutes. And I reviewed the entire workbook. Like, here's the output. Like, it's making all these entries. Like, I'm a 15-year tax pro. I reviewed the work that it did. It looked correct. Like, there's some stuff that the Excel workbook doesn't have, but it completed the workpaper as good as it could have been completed. I actually had to go back and count. It made 349 entries into the spreadsheet from 57 pages of tax docs. Impressive.
But you know what else we tried? What if we had it review the workbook? Like somebody else did the workbook, we had the AI review it. So look at this. I changed the workbook so that there were 10 errors. I made this 401k rollover taxable. I removed a distribution code. I changed a W2 state ID number by one. I removed this 1099 DIV. I changed the year of this state refund. I just made up this 1099 interest. I changed the student loan interest by $20. I misspelled Franchise Tax Board. I added some foreign tax. And I changed an account number. I told it, "Go in grade my paper. Identify any mistakes in here." And in less than a minute, it found the 401k rollover. It found the distribution code I removed. It found the state ID number I changed by one, the 1099 DIV that I removed, the year I changed on the state tax refund. It found your mama's investments that I added, the $20 change to the student loan interest, the French toast board, the 67 account number. But not to worry, it didn't find this $2 that I added to the, uh, foreign tax here on this 1099 DIV. That's a relief. Woo, that thing sucks. So, I asked it, "How did you not catch this?" It apologized profusely, but then said it actually did. It just, it said right here that's an extra value of two in cell D45. So, that was actually my bad. It did. It did find that. This is why, this is why I've genuinely lost sleep. Pretty wild, right?
So, let's talk about what this actually costs to use and will it security? Because this feels like a great time for an ad read. Todd, that's me. I thought you were giving me to look. Oh, I got a brand new accounting ledger to show you. Okay, hear me out. It's a, it's a full accounting ledger like QuickBooks replacement that raised money from OpenAI. OpenAI, company behind ChatGPT, uh, behind maybe the end of the world. Uh-huh. That's right. And it is this company's sponsor, Kick. Hang on. Kick. Kick. It is. It's an accounting ledger that that doesn't feel like it's 10 years old, right? Gosh, the ledgers we use today are long in the teeth. It is self-driving bookkeeping. And gang, they're actually building for accounting firms, which has me so excited. One example of how they understand the stupid pains that we are dealing with right now is Kick handles multi-entity because almost every one of our clients have the main business and then all of these other little side entities, and you're like, I don't want to burn another software subscription on just this dumb holding company or the rental LLC. Yet, the only accounting ledgers that handle multi-entity today are are the huge, super expensive ones. But not Kick. All of a client's entities exist inside of a single workspace. So, they're all in one place. Still like proper separate ledgers, but it means inner-company stuff, it just happens. The platform can see it all. Means your client can see it all in one place. And this is the biggie. You pay for a single subscription for all the entities under one client. So, when their attorney tells them to spin up yet another LLC for their new rental, you are not setting even more money on fire by spinning up yet another QuickBooks file. I know, exciting. And and we are all so ready for a good alternative to QuickBooks and Xero, right? If you are a progressive accounting firm, and if you're this far into a really nerdy AI video, I mean, look in the mirror. Nerd alert. Kick's looking for firms just like yours who join their accountant program and finally have a great alternative to putting more of your clients on QuickBooks and instead having, having a new like AI-first ledger that actually feels like an amazing product in 2026. If you're ready for that, and I, I don't know why you wouldn't be, learn more about Kick down in the video description. Jason, back to you, brother.
So polar bears aside, what does all this cost? So if you haven't used it before, Claude, Claude's just a great online chat assistant like ChatGPT. In the past, it just hasn't been quite as good in my opinion. Claude's online assistant, it's even fine for sensitive client work. In fact, the Claude plan that you would use with your team, it's called the Team Plan. It's similar to ChatGPT for Business. But security and pricing around Claude Co-work, this app that runs locally, looks very different. For one, if you're using Claude Co-work, you will blow through your usage limits on a Team Plan like so fast it's not even funny. Claude will only let you use like so much of their top models. And if you're just on a Team Plan using Co-work, you will get cut off so fast. So the Team Plan, it's $25 per month per user, but with a minimum of five users. And the cap on that plan, it is not enough.
Now, inside your team settings, you can make specific team members premium users, as I've done for myself here. That'll give them a higher usage limit. Beyond that, you can even let them blow through that usage limit if you want. You can see here I'm allowing an additional $200 per month in overages. You can even come down here and set custom spend caps by user. So, there's no ceiling. But if you are using this heavily every day, like let's say you use this as a tax preparer because security aside, it's kind of good enough for that, like this is useful. But if you did, you would probably burn through a couple hundred bucks per week of usage. And in some ways that's a lot. In other ways, maybe it isn't. Well, humans are a lot. And unfortunately, the best plan right now, the best Claude plan, is one of their personal plans, not the team plans that come with a little more security. It's actually, I'll show you, it's the Max plan, which goes up to $200 a month for 20x more usage than Pro. Pro being about what you get on a Team Plan. And I had to use this plan just to put this video together. And I will say the, like, the only way you'll hit the usage limits on this plan is if you have like three or four big tasks always working at the same time. But as with all things AI, this is getting better by the day. In fact, they just released an updated model that's not as smart, but is like a fifth of the cost to run. So today, depending on how you look at it, it may be expensive, but model costs the last few years have gone down like exponentially. And I'd argue if I could do all this at $200 a month securely, and we can't yet, but if this were secure, I'd argue that's an unbelievable value.
But let's talk security. Couple points of clarification here. Claude Co-work runs locally, but the models that are powering it are still running in the cloud. So if I give it a local file to work with, like the model is seeing all that stuff, like it's not as if everything's just happening on this machine. And the solution here, to be very clear, is not local AI models. In my opinion, it never has been the solution, and in the foreseeable future, won't become the solution because you want a model that's very smart, very capable, can see through your mama's investments. In fact, as we're beginning to trust agents to do stuff online on their own, I'm going to demo that next. How smart a model is can be the difference between an agent getting tricked into doing something versus it seeing the risk and avoiding it. It's not dissimilar to phishing attacks in an accounting firm. One person may fall for a phishing attack while the other person sees through it. When you're choosing your AI model, you're going to want the smart model because that comes with real security ramifications. Remember, we're talking about AI agents being left to go out and work autonomously. Like, how smart they are, it matters. And local models, since they're running on your computer, are never going to be as powerful as models that are running in giant server farms.
Now, to be clear, local models are getting really good. Like, they've come so, so far and are perpetually like just lagging a bit behind the very best models that are out there. But when ChatGPT first came out a few years ago, everybody's like, "Well, hold out for the local version." Here we are a few years later, and we still, we just want more and more from the top-of-the-line models, right? So, Co-work runs locally, but all the thinking is happening in the cloud. And while Claude, the chat assistant like ChatGPT, is very secure, comes with, uh, AICPA SOC 2 Type 2 certification. They'll even sign a BAA for HIPAA compliant workflows. That is all very impressive. But Claude Co-work right now, it is in research preview, which means they're still kind of figuring out like where the boundaries are, particularly if you let it use the web. Because if I give it something sensitive, it could potentially share that without my knowledge, right? Let's say there was even a malicious file on my computer in a folder that I had, it was like tasked to do something in, and that file like tricked the agent into doing a no-no. It is all, honestly, uh, exceptionally unlikely. But when a security breach, it's probably the single biggest risk to your accounting firm. And I truly believe this. Like, what else could bring your accounting firm to its knees overnight? Really, security stuff. That's kind of it.
For that reason, for me, it's still too early to trust this with client work. Now, bigger picture, the transition to trusting agents with technical work, brother, it is going to be a messy transition because when will it be the right time? When will it be good enough? And the reality is the details matter. Does it have access to the web or not? Does it have access to my email? Because someone could like email me something malicious that it could get into this thing, Co-work. It is unbelievable. But unfortunately for me, I'm not there yet on sensitive client stuff. In fact, Claude even says probably shouldn't be used for regulated workflows yet. And in the United States, tax work, at least, is definitely a regulated workflow. But you can see like, just how close we are to how powerful this will make not only off-the-shelf software that we buy and use, but how this becomes an incredibly powerful tool to just do my job. Like the job that only I do, that annoying thing that you do for that one client. We're very close to AI just doing that for us, right? And if this has you feeling like scared at all, just real quick, this whole change, like being automated by AI, I'd argue it's already happening for software developers who largely are no longer writing code. They're letting AI write the code. But developers not needing to write code doesn't mean that the world doesn't need developers and accountants. Not needing to account in the traditional sense doesn't mean we won't need accountants. Now, does it mean we'll need, uh, less of them? Quite possibly. But right now, uh, we're a long ways from having an abundance of accounting talent out in the world, right? So things are obviously going to change. We just got to engage in that change. In fact, the most forward-thinking AI accountants in the profession, I'd argue, they're the 1,500 people that are tuning into my podcast each week where we go super deep on this stuff. And we have from the launch of ChatGPT, we were talking about desktop AI agents like 3 years ago. And so, it's just a matter of like learning along the way and being part of it rather than being surprised by it.
We actually just did an episode on the biggest risks that come with AI agents and how to mitigate them. Something we all need to be informed about. In fact, on that subject, let me show you how Claude Co-work can work a browser. And then I'll show you how to swipe all the skills that we've made. So, you can cut Co-work loose in Chrome. It looks like this. And it just requires that you have a Claude Chrome extension connected to your account. Now, this little yellow outline, this is a tab group. And Claude can see any tabs in its group. So, if I log myself into QuickBooks, I've now got this tab and I can drag it into that tab group so that Claude can see it and then come back to Claude and say, "Go run a P&L for me from the last month." And it'll get to work. Very slowly. That is the big headache still with AI agents using the browser. It is slower than molasses on a midwinter Mississippi Monday. To run this P&L, it took about 60 seconds, but I wanted to see if we could trust it with a larger process. So, I downloaded the transactions from the QuickBooks sample file, Craig's Design and Landscaping, and I told my Claude, "Create another 3 months of transaction activity that looks like everything that was in here to date, right? Just make more." I then told Claude, "Pop into my QuickBooks here, add those transactions to QuickBooks and do the bookkeeping." Here's what I did. Sped up at about 20x speed first to try to import the transactions via CSV, but that functionality was blocked because this is a sample company. It then asked how I wanted it to proceed. I said, "Enter it manually via the user interface." So, naturally, it headed to the bank rules screen, then went to the chart of accounts, then stumbled into the register screen for the checking account, and I watched it, tried to put the first of 60-some transactions in there for about 10 minutes, and I decided I couldn't be bothered, pulled the power cord, and hit control. This thing sucks. Now, I almost certainly could have made it faster with like skills or prompts that tell it how I want it to do that, right? But the unfortunate truth right now is that browser use agents, they still suck. Oh, thank goodness. The browser is almost the most dangerous way to use an AI agent because it could stumble into something malicious. I have a four-year-old child. I will happily give it a piece of paper and say, "Do a thing right here, sweet little child. Innocent little baby boy." But I will not go, "Hey, look at the internet. Have a good time, squirt." So, an AI agent working on your files for now, way more useful and probably closer from a security standpoint than trusting something to go out and work on a browser.
But we ran one more test to see if it could go out and attach receipts to the transactions inside of QuickBooks. So, here are 20 receipts that I created with AI. You, you get the skill for that for transactions that are in the QuickBooks sample file. I cut it loose. Look at them go. It immediately went to the expenses screen instead of just searching for the transactions. It worked out how to open a transaction and found the attachments area at the bottom. Good bot. Feels like baby's first accounting file, which it kind of is. And the second time for sure is going to be way more efficient, right? But it, it kept clicking the wrong transaction over and over again. So then it just went and tried search. Here's a bill from the attorney I'm looking for, and whoops, wrong transaction. We wanted the $250 one. So it then typed "Robertson Robertson 250." That didn't give it any results. So it exed out of that. Tried going back to the expenses screen. About 10 minutes had elapsed at that point, and I decided I couldn't be bothered. Sat down here. And now I'm shooting this because this video took like a week and probably a thousand bucks in freaking Claude credits. I need a cigarette, and you might too. Hey honey, no. Uh, I think everything's going to be fine. Fine for now.
I told you I'd show you how to swipe these skills that we made. Like, put a lot of effort into these. So, skills. Skills are just a folder with a specific structure and a markdown file, basically a text file explaining in plain English what that skill can do. Now, you don't even need to know how the structure is supposed to be here because there's a skill for creating skills. Yo, dog, I made a skill so that you'd have the skill of making skills. So, you can literally just come to Claude and say, "Uh, create a skill called 'What Would Jason Say?'" that answers questions about using the thought leadership of major accounting celebrity Jason Stats. I cut it loose. You can see down here on the right, it pulled in its skill creator skill to make the skill. It ended up going out and researching me for a few minutes. We can test it out now. "What would Jason say about finding more profitable clients?" The response looks pretty good. And these skills, they roll up into plugins. So you have like plugins, and under that skills, you also have slash commands. I'm not really a fan. So let's just talk about skills. We can come back here and say now, "Package this into a plugin called 'Useless Tools'." Down here, you can see it's using another skill, the "Create Co-work Plugin" skill. Loved the name. Wow. Okay. And here's the plugin. I've saved it to my Co-work. But for you to install a plugin, you just come over here, click plugins, come in here, click upload.
Now, be aware you could also have like malicious plugins or malicious skills. Imagine something like sending details about you out to some third-party site. Now, that is unrealistic. Let me show you this down here under settings. In fact, you have to enable this one setting for my plugin to work. So, capabilities down here, "Allow network egress." This controls what resources Claude's allowed to go out and use when it needs it. By default, this is turned off. But to like create the PDFs, like the receipts and the bank statements, the stuff that I made the skill for, it needs to go out and grab some additional PDF libraries to do more complicated stuff with those files. There's three levels of security here. Basically, you can have it off entirely. You can have it on for common package managers only. That's probably what I'd recommend. Then it's only going out and like pulling known libraries, like trusted libraries, or you can whitelist additional libraries here, like if you're a real nerd and you have some of your own. Basically, they're just trying to prevent Claude from going out and grabbing something problematic, right? So, if you want to use my skills, make sure that setting's turned on. I will link the skills down below. You literally just have to upload them as a plugin.
But gang, browser use, not good. Everything else, wow. Like, what a time to be alive, gang. Really powerful stuff. And makes you think right now, like using these same AI models, what could you do with an AI agent that just had no restrictions? I could just cut it loose on my computer on the web. Just total Mad Max YOLO mode. How good would that agent be? Well, what you're describing, you might have heard of. Open Claude, ClaudeBot, MoldBot. It's been known by a lot of names, but it was just acquired by OpenAI, and it is a dangerously unlimited AI agent. It's the subject of our next video. Wait till you see all of this. Just working 24 hours a day autonomously on your computer trying to get accounting work done. It is wild. This channel's turning into, uh, stuff you probably shouldn't do. So, let's watch Jason do it instead because someday there will be a secure way to do this. Be sure to subscribe to the channel so you don't miss that one. And if you want to see an AI agent that I think you can actually use today with your accounting firm's client-sensitive data, and my opinion, the best secure agent, it is still ChatGPT's agent mode. Let me show you what those security boundaries look like right now along with some really cool use cases for accounting firms.