Caleb: On today's podcast, we're going to be giving some updates on our use of AI, different ways that we've been coming up with new ideas around AI and new tools that we've been using. ⁓ First, I wanted to say that I've been using N8Analot, that automation platform, and I actually built an AI agent, basically set it up through what's called a platform called Twilio so that I can message it on WhatsApp. And then I gave it tools. So the AI agent itself is just a prompt and Claude. So I give it Claude's API and it's just a prompt. But then I'm building and implementing small tools that it has. Like, for example, it can look at my emails. It can ⁓ create draft replies to my emails. It can delete emails when I ask it. It can read and write. Docs, Google Docs and Google Sheets so it can create new ones, update old ones, get information from Google Docs and Google Sheets that I ask it to. It can. I actually I created a custom app that's like a task app. I think I've talked about this before potentially, but I connected that task app to the agent so that all I have to do is text it like, hey, I need this task put in my task app. Or, you know, what are my top five tasks or something like that? And it pulls all that information and a couple other things. But yeah, it's been pretty cool. I've actually the I will say the coolest part is that I gave it a tool to build and it in tools. So I text the agent on WhatsApp. Hey, can you build this new tool for my agent? And it goes and builds the tool. And all I have to do is connect my credentials and connect it to the agent. but the tool's built by the agent itself. So it's pretty cool, yeah. That's one of the biggest things I've been working on recently. A couple other things that I'll talk about, but that has been very helpful and I can just talk to it while I'm out and not even at my computer. And I also gave it access to cloud code. So it's been crazy, the stuff that it can do. It's pretty wild.
J&C Unscripted: ⁓ What main functions are you using it for? Mostly just like task tracking, calendar stuff.
Caleb: Yeah, so I gave it access to email, Google Drive, which is the Google Docs and Sheets, Google Calendar, my task app, N8N, Claude code. And the main one I use it for is. Honestly, I'll come up with an idea, talk through it with Claude, through the agent, and then have it save it as a Google account or I mean, save it as a Google Doc. or add an event to my calendar or add a task. And then if I think of a new tool idea, then I use N8n, the tool creator. I haven't used the email one as much because I'm trying to get used to it a little more. It's weird not like looking at the emails and just having an agent tell you what your emails are. But I genuinely think it would be helpful because then I wouldn't have to actually log in to email every time. The main point is not really
J&C Unscripted: Mm-hmm.
Caleb: The main point for me is not having to log into all this stuff. So I don't have to log into email. I don't have to log into Google Drive. I don't have to log into NNN. I don't have to log into Cloud Code. It just does it for me because it's connected to an API. So that's the biggest thing is saving me time from logging into everything.
J&C Unscripted: I take it you don't just stay logged into all those things.
Caleb: Well, I do for most of those things, but I guess when I say logging in, not just like username, password, login, but opening the tab, finding what I need, making the update instead of just like, hey, do this for me. And then it does it.
J&C Unscripted: like opening. Mm-hmm. Yeah, I I still have yet to find ⁓ a daily use for AI outside of the realm of ⁓ Search like a more a smarter search function ⁓ Because when I think about it, it's like well, yeah, I could integrate it great with some task tracking But I do all that mostly manually now. So
Caleb: Yeah.
J&C Unscripted: Doesn't matter. ⁓ So I'm always on the search for something that, until I have a robot, I think it's somewhat limited though.
Caleb: Yeah. So actually one thing that I'm adding, and I'm pulling it up right now, but one thing I'm trying to add is a QuickBooks transaction categorizer. So basically I message it, say, what are my last 10 uncategorized transactions? And then it tells me what they are, gives me a recommendation for the transaction, and I can either say like, yes or no, here's the actual category.
J&C Unscripted: Hmm
Caleb: I think that one is going to be a big one because I never feel like logging into QuickBooks. Like that one I actually have to log into every single time and I'm not always at my computer and so sometimes I get behind because I try to do personal accounts on QuickBooks. I definitely do not use the account in QuickBooks for all my personal stuff.
J&C Unscripted: Ha Uh-huh. Yeah.
Caleb: No, I definitely do do that. So it just takes a while. Like if I'm not on my computer, then I forget that I don't log in. Then I go a week without categorizing. Then it just gets annoying because there's too many like there gets to be a lot of transactions. So I think that one's going to save me mental space because I'll just ask what my transactions are and it will categorize it.
J&C Unscripted: Mm-hmm. Yeah. Right.
Caleb: That one could potentially be useful if even even if you connected it to honestly, I'm trying to think of like what a business owner find that useful or would it be more? I don't know, but for me, it's at least useful because I do have I know what I want.
J&C Unscripted: The business owner would probably find it useful. The accountant won't for a while, but ⁓ at Digits, which is AI bookkeeping, they did a test that was like categorizing. It was several hundred, I believe, transactions, and they benchmarked it versus a group of bookkeepers. And I think it was... Like they already knew what the correct category should be for each of the transactions. ⁓ And the AI through their bookkeeping software, I think it was like it got 80, I don't know, was like 80 % of everything correct. And then the bookkeepers was like 85. So, you know, we're approaching singularity, but ⁓ man, QuickBooks tools are terrible.
Caleb: Yeah. So actually that brings up something I didn't even think about talking about, but we've had this discussion a lot about quantity versus quality, like efficiency, that kind of discussion. But I'm not sure that we've ever had the conversation about what happens when, and I don't even know what this would be called, but. over time. So let's say a new technology increases our speed but decreases our quality by 5%. Most of the time we would probably be okay with that. Like if it 10x'd our speed, but then decreased our quality by 5%. However, then there's another technology. So we're already 5 % lower. This technology increases our speed by 5x or 10x, but then decreases our quality by 5%. Then there's a new technology that increases our speed by, so like over time it's like this 5 % becomes.
J&C Unscripted: Mm-hmm.
Caleb: 10 % becomes 15 % becomes 20 for all the new technological revolutions. How would you think about still balancing speed and quality if over time our quality is being degraded because we are constantly increasing our speed and giving up a little bit of quality each time? Like how would you think about that?
J&C Unscripted: Yeah, in the accounting profession, which is more like what I could speak to, there's definitely, we've had it with some clients where it's like the, there's a certain transactions under a certain value are meaningless unless they happen all the time. So that falls into your margin of error quite fine because there's this huge pool of $5, $10, $15 transactions that reoccur that are all getting an expense, going into an expense column and they don't change your books that much to change how you should make business decisions. ⁓ And so it doesn't matter. inaccuracy in that area, as long as it hits an expense item is not meaningful. ⁓ So that's error we just don't care about. ⁓ The other thought I had while you were saying that is I was like, ⁓ well, this is control theory ⁓ Which the controls theory Like it kind of gets back to ⁓ Claude Shannon and the making of computers and all that which is where you're integrating a feedback loop into your systems So I really think that going forward we're going to see more AIs ⁓ doing a review function, which is also what I want to have access to for books because having an AI do work, not on your time scale that can run at night, that can review the entire general ledger and be like, what's this? What's this? That was odd. Hey, last time you had something, ⁓ Last time a client emailed you about disposing a thing, you moved it from equipment to disposed equipment. But I noticed you haven't done that yet. And we got an email last week from the client, ⁓ like that kind of thing, ⁓ which that works into kind of controls. So you can still use technology to increase your speed. But then, so like in the bookkeeping case, it would be the AR categorizes during the month.
Caleb: Yeah.
J&C Unscripted: Meanwhile, you've got a second eye at AI. All it's doing is overlord. It's just watching everything and monitoring your email, your communications, and the books, and making sure those make sense and anything that's odd flags for human review. So you can still increase efficiency and speed and ⁓ focus on error correction.
Caleb: So I think that's possible right now even, like having, creating a review AI. The only problem would be that I know you guys text a lot with business owners and so having access to the text would be hard unless there were specific phones. Like if people use iPhones and stuff, Apple is super dumb about that. But.
J&C Unscripted: Mm-hmm.
Caleb: It's whatever. What would you say are like the biggest if you were to implement an AI system that did some things for you? What features would it need? To make it work. At work.
J&C Unscripted: For me personally or me at work? Yeah, I need to be absolutely sure. First of all, nobody's data is going anywhere else. Like it's gotta be an in-house model or a local AI ⁓ and nothing is going to get out into the broader models. Obviously that's number one Beyond that it's like I I want to be able to have a ⁓ An intern not even an intern ⁓ a green employee a somewhat green employee Who is more diligent than a great employee That ⁓ a books reviewer would be ideal
Caleb: Yeah
J&C Unscripted: ⁓ And what that would entail is I need this thing to be able to pull in context Not only from this QuickBooks file But I need it to be able to say here are all the other clients we have that are like this client Go look at them. I need this QuickBooks file to be like that ⁓
Caleb: If it was a new client, you're Like if you're setting up a new client?
J&C Unscripted: Yeah, if it was a new client or even just in training it, like this is a client that is in this field, they use properties. This is a client that has rental properties or flips. We use classes to differentiate, blah, blah, blah, et cetera, et cetera. Here are the clients we already do that for. Train yourself on our modus operandi for clients with rental properties, that kind of thing. ⁓ That way I could have a reviewer and I could have a repository of knowledge that ⁓ if I'm on vacation and somebody else is doing the books for me, they can easily ask questions for from. ⁓
Caleb: Yeah, yeah, that makes sense. Yeah.
J&C Unscripted: And then beyond that, it would be ⁓ the next step for having it generate the documents for an MCP, just like you would have an associate generate them for a manager, would be really easy. Once you've got the full context of the books, then it's just executing the run these reports and put them in this folder. ⁓ And then you start getting into individual tasks. Do I trust it to categorize books completely? Well, I'll know that after I've used it for a while to run an MCP. Like what do you catch and what do you don't catch? And then you're getting into the like special reporting, draft me annual meeting minutes. ⁓ or draft meet PPT for the year, et cetera, et cetera.
Caleb: Yeah, I wonder if it could draft. Well, I mean, you use it. Specific set of documents to draft PPT, so I wonder if you could just give it those documents and it could draft it for you. That would be really. Yeah, no, that's a. It would be cool if it could. ⁓
J&C Unscripted: Mm-hmm.
Caleb: if you could build something that monitored communications so that you it could see what the business owner is asking you to do and then just go do it. I'm sure that it's possible, but that would be, that'd be pretty cool.
J&C Unscripted: Mm-hmm. I think we're close. I think even by doing nothing within 24 months. ⁓
Caleb: it will be in QuickBooks most likely. There's a QuickBooks. It's like QuickBooks intelligence beta or something, which I'm kind of skeptical because QuickBooks is never good at AI. However, there's also a QuickBooks MCP, which is like ⁓ basically a connector that it's like a...
J&C Unscripted: And it's terrible. It's awful. Hmm.
Caleb: you can connect AI systems to QuickBooks and ask questions. So it might be possible to do at least part of what you're saying just from that, where you can get context from other QuickBooks files for a new file just because you can ask it questions and stuff. You should check it out.
J&C Unscripted: Yeah. I think that's eventually how it's gonna happen. I don't think Intuit's gonna do it, because it's terrible. Their AI is awful. Yes. Right. It's just, it's egregious. Yeah, I think somebody else will make a plugin of some kind and integrate with the API. ⁓ It'll be like the right tool does. That's what I think. But it's good enough.
Caleb: They need to just stay focused on accounting and bookkeeping and not on the AI stuff. you Yeah.
J&C Unscripted: I think somebody could do it. It just, and I'm sure somebody's working on it, but it seems like people who are working on it are working on their own bookkeeping ⁓ ledgers. There's digits and there's, think it's, does it count? There's another one.
Caleb: Yeah. Have you tried digits? Like have you tried ⁓ actually digits?
J&C Unscripted: Mm-hmm. No, I listened to an interview with the CEO and ⁓ My guess is that it's gonna be most useful for Like individual business owners right now, but I could be wrong ⁓ He said it still doesn't get like splitting transactions very well ⁓ Like categorized as well, but that's kind of where the limitation is currently ⁓ So it's not quite there yet, but yeah, I think fundamentally bookkeeping changes in two years. And I welcome it.
Caleb: Yeah. Yeah, that makes sense. So I know I got off on a really a big tangent, but going back to the whole control theory thing, ⁓ what, so I know you talked about having AI that does review. Do you think that there is a bit of, I don't know, bias or? I'm not sure necessarily how to call it or what to call it, but if an AI, if the same AI system is doing the review as a system that is doing the initial categorization or tasks, wouldn't there be some sort of blind spot that it wouldn't be able to catch at least some of the errors that are made because it is the same system and it's trained on the same data?
J&C Unscripted: you Well, yeah, you could have it be the same system or separate. I mean, you could use a different model or two if you wanted. Like the nice part about autonomous agents, ⁓ like I was looking into perplexity computer over the weekend ⁓ last week, but it doesn't seem to be where I want it to be yet. But that's fine. It's a great step in a good direction. ⁓ But with autonomous agents when you run your MCP you could train two entirely separate models and ⁓ Run both run it through one and then run it through the other ⁓ And see what one catches and one doesn't ⁓ Like that that's the beauty of it and since it's asynchronous it can happen while nobody's work
Caleb: Yeah.
J&C Unscripted: just run this thing. It doesn't matter if it takes 24 hours because you didn't have to do it.
Caleb: Yeah, no, that makes sense.
J&C Unscripted: There's like one circuitry we used in ⁓ engineering, which is especially relevant for wireless transmission, ⁓ like radio signals and whatever, is that those signals are so small, you have to have an amplifier that increases them by a large degree. But an amplifier that amplifies a large size and is of high quality and low margin rate is quite expensive. So what you can do is you find an amplifier that doesn't amplify much at all ⁓ but only has two qualities. It's fast, very fast and very precise but doesn't amplify much at all. You then take whatever your crappiest cheapest amplifier that is and you can put strings of them 5 6 7 8 9 10 string them together and but it's not accurate but it has the ability to amplify you know 100x or 1000x or whatever you then take your fast and precise amplifier and you put that string of crappy ones into the feedback loop of it. So you tie the feedback symbol input of the good amplifier to the output of your last crappy amplifier. And as long as this guy is, I believe roughly 10 times faster than these crappy amplifiers, you can get a extremely high amplification at no error rate, like practically nothing, because this crappy amplifier is controlling that whole thing. ⁓ So that's kind of like control theory when it goes well. So you're getting the best of the two components ⁓ and for a cheaper cost than it would be to have just bought one really nice one.
Caleb: Yeah. Huh. Interesting. Yeah, I've just I didn't even really thought about that until just now when I was like, well, if we give up some quality multiple times, then we're giving up like exponential quality and not just five total percent. But I do think there's ways like you give up some quality and then the quality of the. technology gets better. So then you get gained some of the quality back. Then you give up a little bit and then it gets better. You gain some of it back on and on and on. So I don't necessarily go ahead.
J&C Unscripted: Mm-hmm and and in in our field there's a lot that is missed not because of competence But because at the end of an eight hour day a human can only look over so many transactions And that's the kind of thing where like those things will actually increase in accuracy And it's not because of any any competence of the bookkeeper it's because humans can just do so much
Caleb: Yeah. Yeah, I wonder what the the maximum amount of transaction someone could categorize in an eight hour work day is. Five thousand. No, that'd be insane.
J&C Unscripted: That sounds terrible. It probably depends on if they're all of similar vendors or all different. Because anything for Wendy's is going into meals.
Caleb: Yeah, with quick looks. Yeah, true. That would be really interesting. But also really boring. So I don't think anyone would ever want to do it.
J&C Unscripted: You Exactly why I want an AI. ⁓
Caleb: Yeah. Yeah, I think there's a lot of uses for AI and bookkeeping. It's just, and I'm sure there's a lot of bookkeeping softwares that are implementing AI. It's just finding a dominant player that's actually good because QuickBooks is not.
J&C Unscripted: Mm-hmm the the Where I've seen AI work in our workflows which we're using them very limited but ⁓ have been in tasks that are ⁓ Well, that's not true. There's some reoccurring but not a ton of reoccurring. It's kind of in the one-off ⁓ I need to research a question or I Need this specialty thing created. So the one we had happen this week is a client is Applying for a loan For a vehicle that he wants to go look at but he doesn't know if he wants to purchase this purchase it until he looks at it but it's hours away. So he wants to be ready to buy it when he goes to look at it. ⁓ Well, the bank wants a bill of sale or like a, not a contract, but I can't remember how they termed it. The client doesn't want to get signatures on a bill of sale for a car they're not buying until they look at. ⁓
Caleb: Yeah.
J&C Unscripted: So my coworker basically outlined the scenario and tried to give it to perplexity. Perplexity was way off. It created like this gigantic legalese form ⁓ that just didn't make any sense. And so he then threw it into Claude and then had to do a little bit of coaching. ⁓ But I didn't have any client information. Like you just said, here's the description. I need a bill of sale. ⁓ This contingent on like this is an agreement to purchase the thing the buyer is going to review the object, the car, vehicle. And if he doesn't like it, he's not legally required to purchase it. And then I had a place for the owner to sign, the owner and the seller. And the bank accepted that. ⁓
Caleb: Hey, Claude's better. Claude's better.
J&C Unscripted: Yeah, I still like perplexity for search and research but um My experience with perplexity is like I also did this week. I had an excel sheet That had no client info on it threw it into perplexity. It was like I'm kind of bad at making these things look visually appealing. Can you clean it up for me? It was terrible. It added like 20 columns
Caleb: Yeah, for research, definitely.
J&C Unscripted: and then tried to make all these descriptions that like was not concise. ⁓ So that was a failure. Now I've done a terrible job of testing Claude.
Caleb: Did you throw it in clod and try it in clod? You should. I will say, so I would say chat GPT is best to images. Claude is best at writing and coding and perplexity is best at research. So if you have, I would say analysis is also kind of in the whole coding area with Claude. So try it, try it in Claude for sure.
J&C Unscripted: Mmm. Alright, I'll try that. I did have a big win with perplexity though. youngest needs an echocardiogram and so I've been searching around for providers and me and perplexity have a space where I ask you questions and I finally got it to spit out a very good ⁓ comprehensive list of providers in the area that may offer The problem is pediatric ⁓ echocardiogram. That's what I need. ⁓ so it gave me a good list to run with. And it even gave me a script. Here's how the conversation should go. Well, what I was impressed with is its list of things to ask about, which I know from past experience ⁓ was, do the doctor
Caleb: Yeah.
J&C Unscripted: any of the technician or facility or the results reading bill separately. And I talked to the Riley in ⁓ Indianapolis. I talked to the estimate department at Riley. They didn't know if they send echocardiogram results to a third party to read, which will sometimes bill separately. Yeah, yeah.
Caleb: that how do you not know?
J&C Unscripted: So then they send me to the billing department where you pay. They said maybe they would know. They didn't Although at least that lady was helpful and she thought she knew the answer. And I was like, do I just need to call the facility and ask? And she's like, yeah, I would do that. So I called the receptionist at the testing facility who then had to walk in the back and ask somebody if they send results out.
Caleb: That's crazy. but false confidence.
J&C Unscripted: to get red. I know isn't that nuts? No one knows. Yeah, yeah. So good job perplexity.
Caleb: That's crazy. No one has any, the process is, wow. That's fun. Yeah, no, mean, perplexity is definitely the best research and you pay for pro, right?
J&C Unscripted: Yeah, I've got it free for a year, but yeah
Caleb: ⁓ yeah, from the Venmo thing. Nice. Yep. It's really good at research. ⁓ You got anything else?
J&C Unscripted: I don't think so, but we'll keep you updated when AI replaces us on this podcast.
Caleb: Who says already hasn't? No, I'm just kidding. All right, well, that's all I have to. So ⁓ make sure you guys like and subscribe on YouTube. Five star review on Spotify if you enjoy the episode. And we'd love to hear from you on email at jc.unscripted at gmail.com. That is the pod, and we will see you on the next one.
J&C Unscripted: ⁓ What main functions are you using it for? Mostly just like task tracking, calendar stuff.
Caleb: Yeah, so I gave it access to email, Google Drive, which is the Google Docs and Sheets, Google Calendar, my task app, N8N, Claude code. And the main one I use it for is. Honestly, I'll come up with an idea, talk through it with Claude, through the agent, and then have it save it as a Google account or I mean, save it as a Google Doc. or add an event to my calendar or add a task. And then if I think of a new tool idea, then I use N8n, the tool creator. I haven't used the email one as much because I'm trying to get used to it a little more. It's weird not like looking at the emails and just having an agent tell you what your emails are. But I genuinely think it would be helpful because then I wouldn't have to actually log in to email every time. The main point is not really
J&C Unscripted: Mm-hmm.
Caleb: The main point for me is not having to log into all this stuff. So I don't have to log into email. I don't have to log into Google Drive. I don't have to log into NNN. I don't have to log into Cloud Code. It just does it for me because it's connected to an API. So that's the biggest thing is saving me time from logging into everything.
J&C Unscripted: I take it you don't just stay logged into all those things.
Caleb: Well, I do for most of those things, but I guess when I say logging in, not just like username, password, login, but opening the tab, finding what I need, making the update instead of just like, hey, do this for me. And then it does it.
J&C Unscripted: like opening. Mm-hmm. Yeah, I I still have yet to find ⁓ a daily use for AI outside of the realm of ⁓ Search like a more a smarter search function ⁓ Because when I think about it, it's like well, yeah, I could integrate it great with some task tracking But I do all that mostly manually now. So
Caleb: Yeah.
J&C Unscripted: Doesn't matter. ⁓ So I'm always on the search for something that, until I have a robot, I think it's somewhat limited though.
Caleb: Yeah. So actually one thing that I'm adding, and I'm pulling it up right now, but one thing I'm trying to add is a QuickBooks transaction categorizer. So basically I message it, say, what are my last 10 uncategorized transactions? And then it tells me what they are, gives me a recommendation for the transaction, and I can either say like, yes or no, here's the actual category.
J&C Unscripted: Hmm
Caleb: I think that one is going to be a big one because I never feel like logging into QuickBooks. Like that one I actually have to log into every single time and I'm not always at my computer and so sometimes I get behind because I try to do personal accounts on QuickBooks. I definitely do not use the account in QuickBooks for all my personal stuff.
J&C Unscripted: Ha Uh-huh. Yeah.
Caleb: No, I definitely do do that. So it just takes a while. Like if I'm not on my computer, then I forget that I don't log in. Then I go a week without categorizing. Then it just gets annoying because there's too many like there gets to be a lot of transactions. So I think that one's going to save me mental space because I'll just ask what my transactions are and it will categorize it.
J&C Unscripted: Mm-hmm. Yeah. Right.
Caleb: That one could potentially be useful if even even if you connected it to honestly, I'm trying to think of like what a business owner find that useful or would it be more? I don't know, but for me, it's at least useful because I do have I know what I want.
J&C Unscripted: The business owner would probably find it useful. The accountant won't for a while, but ⁓ at Digits, which is AI bookkeeping, they did a test that was like categorizing. It was several hundred, I believe, transactions, and they benchmarked it versus a group of bookkeepers. And I think it was... Like they already knew what the correct category should be for each of the transactions. ⁓ And the AI through their bookkeeping software, I think it was like it got 80, I don't know, was like 80 % of everything correct. And then the bookkeepers was like 85. So, you know, we're approaching singularity, but ⁓ man, QuickBooks tools are terrible.
Caleb: Yeah. So actually that brings up something I didn't even think about talking about, but we've had this discussion a lot about quantity versus quality, like efficiency, that kind of discussion. But I'm not sure that we've ever had the conversation about what happens when, and I don't even know what this would be called, but. over time. So let's say a new technology increases our speed but decreases our quality by 5%. Most of the time we would probably be okay with that. Like if it 10x'd our speed, but then decreased our quality by 5%. However, then there's another technology. So we're already 5 % lower. This technology increases our speed by 5x or 10x, but then decreases our quality by 5%. Then there's a new technology that increases our speed by, so like over time it's like this 5 % becomes.
J&C Unscripted: Mm-hmm.
Caleb: 10 % becomes 15 % becomes 20 for all the new technological revolutions. How would you think about still balancing speed and quality if over time our quality is being degraded because we are constantly increasing our speed and giving up a little bit of quality each time? Like how would you think about that?
J&C Unscripted: Yeah, in the accounting profession, which is more like what I could speak to, there's definitely, we've had it with some clients where it's like the, there's a certain transactions under a certain value are meaningless unless they happen all the time. So that falls into your margin of error quite fine because there's this huge pool of $5, $10, $15 transactions that reoccur that are all getting an expense, going into an expense column and they don't change your books that much to change how you should make business decisions. ⁓ And so it doesn't matter. inaccuracy in that area, as long as it hits an expense item is not meaningful. ⁓ So that's error we just don't care about. ⁓ The other thought I had while you were saying that is I was like, ⁓ well, this is control theory ⁓ Which the controls theory Like it kind of gets back to ⁓ Claude Shannon and the making of computers and all that which is where you're integrating a feedback loop into your systems So I really think that going forward we're going to see more AIs ⁓ doing a review function, which is also what I want to have access to for books because having an AI do work, not on your time scale that can run at night, that can review the entire general ledger and be like, what's this? What's this? That was odd. Hey, last time you had something, ⁓ Last time a client emailed you about disposing a thing, you moved it from equipment to disposed equipment. But I noticed you haven't done that yet. And we got an email last week from the client, ⁓ like that kind of thing, ⁓ which that works into kind of controls. So you can still use technology to increase your speed. But then, so like in the bookkeeping case, it would be the AR categorizes during the month.
Caleb: Yeah.
J&C Unscripted: Meanwhile, you've got a second eye at AI. All it's doing is overlord. It's just watching everything and monitoring your email, your communications, and the books, and making sure those make sense and anything that's odd flags for human review. So you can still increase efficiency and speed and ⁓ focus on error correction.
Caleb: So I think that's possible right now even, like having, creating a review AI. The only problem would be that I know you guys text a lot with business owners and so having access to the text would be hard unless there were specific phones. Like if people use iPhones and stuff, Apple is super dumb about that. But.
J&C Unscripted: Mm-hmm.
Caleb: It's whatever. What would you say are like the biggest if you were to implement an AI system that did some things for you? What features would it need? To make it work. At work.
J&C Unscripted: For me personally or me at work? Yeah, I need to be absolutely sure. First of all, nobody's data is going anywhere else. Like it's gotta be an in-house model or a local AI ⁓ and nothing is going to get out into the broader models. Obviously that's number one Beyond that it's like I I want to be able to have a ⁓ An intern not even an intern ⁓ a green employee a somewhat green employee Who is more diligent than a great employee That ⁓ a books reviewer would be ideal
Caleb: Yeah
J&C Unscripted: ⁓ And what that would entail is I need this thing to be able to pull in context Not only from this QuickBooks file But I need it to be able to say here are all the other clients we have that are like this client Go look at them. I need this QuickBooks file to be like that ⁓
Caleb: If it was a new client, you're Like if you're setting up a new client?
J&C Unscripted: Yeah, if it was a new client or even just in training it, like this is a client that is in this field, they use properties. This is a client that has rental properties or flips. We use classes to differentiate, blah, blah, blah, et cetera, et cetera. Here are the clients we already do that for. Train yourself on our modus operandi for clients with rental properties, that kind of thing. ⁓ That way I could have a reviewer and I could have a repository of knowledge that ⁓ if I'm on vacation and somebody else is doing the books for me, they can easily ask questions for from. ⁓
Caleb: Yeah, yeah, that makes sense. Yeah.
J&C Unscripted: And then beyond that, it would be ⁓ the next step for having it generate the documents for an MCP, just like you would have an associate generate them for a manager, would be really easy. Once you've got the full context of the books, then it's just executing the run these reports and put them in this folder. ⁓ And then you start getting into individual tasks. Do I trust it to categorize books completely? Well, I'll know that after I've used it for a while to run an MCP. Like what do you catch and what do you don't catch? And then you're getting into the like special reporting, draft me annual meeting minutes. ⁓ or draft meet PPT for the year, et cetera, et cetera.
Caleb: Yeah, I wonder if it could draft. Well, I mean, you use it. Specific set of documents to draft PPT, so I wonder if you could just give it those documents and it could draft it for you. That would be really. Yeah, no, that's a. It would be cool if it could. ⁓
J&C Unscripted: Mm-hmm.
Caleb: if you could build something that monitored communications so that you it could see what the business owner is asking you to do and then just go do it. I'm sure that it's possible, but that would be, that'd be pretty cool.
J&C Unscripted: Mm-hmm. I think we're close. I think even by doing nothing within 24 months. ⁓
Caleb: it will be in QuickBooks most likely. There's a QuickBooks. It's like QuickBooks intelligence beta or something, which I'm kind of skeptical because QuickBooks is never good at AI. However, there's also a QuickBooks MCP, which is like ⁓ basically a connector that it's like a...
J&C Unscripted: And it's terrible. It's awful. Hmm.
Caleb: you can connect AI systems to QuickBooks and ask questions. So it might be possible to do at least part of what you're saying just from that, where you can get context from other QuickBooks files for a new file just because you can ask it questions and stuff. You should check it out.
J&C Unscripted: Yeah. I think that's eventually how it's gonna happen. I don't think Intuit's gonna do it, because it's terrible. Their AI is awful. Yes. Right. It's just, it's egregious. Yeah, I think somebody else will make a plugin of some kind and integrate with the API. ⁓ It'll be like the right tool does. That's what I think. But it's good enough.
Caleb: They need to just stay focused on accounting and bookkeeping and not on the AI stuff. you Yeah.
J&C Unscripted: I think somebody could do it. It just, and I'm sure somebody's working on it, but it seems like people who are working on it are working on their own bookkeeping ⁓ ledgers. There's digits and there's, think it's, does it count? There's another one.
Caleb: Yeah. Have you tried digits? Like have you tried ⁓ actually digits?
J&C Unscripted: Mm-hmm. No, I listened to an interview with the CEO and ⁓ My guess is that it's gonna be most useful for Like individual business owners right now, but I could be wrong ⁓ He said it still doesn't get like splitting transactions very well ⁓ Like categorized as well, but that's kind of where the limitation is currently ⁓ So it's not quite there yet, but yeah, I think fundamentally bookkeeping changes in two years. And I welcome it.
Caleb: Yeah. Yeah, that makes sense. So I know I got off on a really a big tangent, but going back to the whole control theory thing, ⁓ what, so I know you talked about having AI that does review. Do you think that there is a bit of, I don't know, bias or? I'm not sure necessarily how to call it or what to call it, but if an AI, if the same AI system is doing the review as a system that is doing the initial categorization or tasks, wouldn't there be some sort of blind spot that it wouldn't be able to catch at least some of the errors that are made because it is the same system and it's trained on the same data?
J&C Unscripted: you Well, yeah, you could have it be the same system or separate. I mean, you could use a different model or two if you wanted. Like the nice part about autonomous agents, ⁓ like I was looking into perplexity computer over the weekend ⁓ last week, but it doesn't seem to be where I want it to be yet. But that's fine. It's a great step in a good direction. ⁓ But with autonomous agents when you run your MCP you could train two entirely separate models and ⁓ Run both run it through one and then run it through the other ⁓ And see what one catches and one doesn't ⁓ Like that that's the beauty of it and since it's asynchronous it can happen while nobody's work
Caleb: Yeah.
J&C Unscripted: just run this thing. It doesn't matter if it takes 24 hours because you didn't have to do it.
Caleb: Yeah, no, that makes sense.
J&C Unscripted: There's like one circuitry we used in ⁓ engineering, which is especially relevant for wireless transmission, ⁓ like radio signals and whatever, is that those signals are so small, you have to have an amplifier that increases them by a large degree. But an amplifier that amplifies a large size and is of high quality and low margin rate is quite expensive. So what you can do is you find an amplifier that doesn't amplify much at all ⁓ but only has two qualities. It's fast, very fast and very precise but doesn't amplify much at all. You then take whatever your crappiest cheapest amplifier that is and you can put strings of them 5 6 7 8 9 10 string them together and but it's not accurate but it has the ability to amplify you know 100x or 1000x or whatever you then take your fast and precise amplifier and you put that string of crappy ones into the feedback loop of it. So you tie the feedback symbol input of the good amplifier to the output of your last crappy amplifier. And as long as this guy is, I believe roughly 10 times faster than these crappy amplifiers, you can get a extremely high amplification at no error rate, like practically nothing, because this crappy amplifier is controlling that whole thing. ⁓ So that's kind of like control theory when it goes well. So you're getting the best of the two components ⁓ and for a cheaper cost than it would be to have just bought one really nice one.
Caleb: Yeah. Huh. Interesting. Yeah, I've just I didn't even really thought about that until just now when I was like, well, if we give up some quality multiple times, then we're giving up like exponential quality and not just five total percent. But I do think there's ways like you give up some quality and then the quality of the. technology gets better. So then you get gained some of the quality back. Then you give up a little bit and then it gets better. You gain some of it back on and on and on. So I don't necessarily go ahead.
J&C Unscripted: Mm-hmm and and in in our field there's a lot that is missed not because of competence But because at the end of an eight hour day a human can only look over so many transactions And that's the kind of thing where like those things will actually increase in accuracy And it's not because of any any competence of the bookkeeper it's because humans can just do so much
Caleb: Yeah. Yeah, I wonder what the the maximum amount of transaction someone could categorize in an eight hour work day is. Five thousand. No, that'd be insane.
J&C Unscripted: That sounds terrible. It probably depends on if they're all of similar vendors or all different. Because anything for Wendy's is going into meals.
Caleb: Yeah, with quick looks. Yeah, true. That would be really interesting. But also really boring. So I don't think anyone would ever want to do it.
J&C Unscripted: You Exactly why I want an AI. ⁓
Caleb: Yeah. Yeah, I think there's a lot of uses for AI and bookkeeping. It's just, and I'm sure there's a lot of bookkeeping softwares that are implementing AI. It's just finding a dominant player that's actually good because QuickBooks is not.
J&C Unscripted: Mm-hmm the the Where I've seen AI work in our workflows which we're using them very limited but ⁓ have been in tasks that are ⁓ Well, that's not true. There's some reoccurring but not a ton of reoccurring. It's kind of in the one-off ⁓ I need to research a question or I Need this specialty thing created. So the one we had happen this week is a client is Applying for a loan For a vehicle that he wants to go look at but he doesn't know if he wants to purchase this purchase it until he looks at it but it's hours away. So he wants to be ready to buy it when he goes to look at it. ⁓ Well, the bank wants a bill of sale or like a, not a contract, but I can't remember how they termed it. The client doesn't want to get signatures on a bill of sale for a car they're not buying until they look at. ⁓
Caleb: Yeah.
J&C Unscripted: So my coworker basically outlined the scenario and tried to give it to perplexity. Perplexity was way off. It created like this gigantic legalese form ⁓ that just didn't make any sense. And so he then threw it into Claude and then had to do a little bit of coaching. ⁓ But I didn't have any client information. Like you just said, here's the description. I need a bill of sale. ⁓ This contingent on like this is an agreement to purchase the thing the buyer is going to review the object, the car, vehicle. And if he doesn't like it, he's not legally required to purchase it. And then I had a place for the owner to sign, the owner and the seller. And the bank accepted that. ⁓
Caleb: Hey, Claude's better. Claude's better.
J&C Unscripted: Yeah, I still like perplexity for search and research but um My experience with perplexity is like I also did this week. I had an excel sheet That had no client info on it threw it into perplexity. It was like I'm kind of bad at making these things look visually appealing. Can you clean it up for me? It was terrible. It added like 20 columns
Caleb: Yeah, for research, definitely.
J&C Unscripted: and then tried to make all these descriptions that like was not concise. ⁓ So that was a failure. Now I've done a terrible job of testing Claude.
Caleb: Did you throw it in clod and try it in clod? You should. I will say, so I would say chat GPT is best to images. Claude is best at writing and coding and perplexity is best at research. So if you have, I would say analysis is also kind of in the whole coding area with Claude. So try it, try it in Claude for sure.
J&C Unscripted: Mmm. Alright, I'll try that. I did have a big win with perplexity though. youngest needs an echocardiogram and so I've been searching around for providers and me and perplexity have a space where I ask you questions and I finally got it to spit out a very good ⁓ comprehensive list of providers in the area that may offer The problem is pediatric ⁓ echocardiogram. That's what I need. ⁓ so it gave me a good list to run with. And it even gave me a script. Here's how the conversation should go. Well, what I was impressed with is its list of things to ask about, which I know from past experience ⁓ was, do the doctor
Caleb: Yeah.
J&C Unscripted: any of the technician or facility or the results reading bill separately. And I talked to the Riley in ⁓ Indianapolis. I talked to the estimate department at Riley. They didn't know if they send echocardiogram results to a third party to read, which will sometimes bill separately. Yeah, yeah.
Caleb: that how do you not know?
J&C Unscripted: So then they send me to the billing department where you pay. They said maybe they would know. They didn't Although at least that lady was helpful and she thought she knew the answer. And I was like, do I just need to call the facility and ask? And she's like, yeah, I would do that. So I called the receptionist at the testing facility who then had to walk in the back and ask somebody if they send results out.
Caleb: That's crazy. but false confidence.
J&C Unscripted: to get red. I know isn't that nuts? No one knows. Yeah, yeah. So good job perplexity.
Caleb: That's crazy. No one has any, the process is, wow. That's fun. Yeah, no, mean, perplexity is definitely the best research and you pay for pro, right?
J&C Unscripted: Yeah, I've got it free for a year, but yeah
Caleb: ⁓ yeah, from the Venmo thing. Nice. Yep. It's really good at research. ⁓ You got anything else?
J&C Unscripted: I don't think so, but we'll keep you updated when AI replaces us on this podcast.
Caleb: Who says already hasn't? No, I'm just kidding. All right, well, that's all I have to. So ⁓ make sure you guys like and subscribe on YouTube. Five star review on Spotify if you enjoy the episode. And we'd love to hear from you on email at jc.unscripted at gmail.com. That is the pod, and we will see you on the next one.
