In this video, the presenter demonstrates a live experiment to categorize over 10,000 synthetic bank transactions for an accounting‑firm use case using a low‑cost, choice‑only AI model called Jeb from Typesafe. He first explains the accounting‑firm problem: premium models (Claude, GPT‑4) become prohibitively expensive when run on every transaction, so firms need a cheaper alternative. He generates 10,000 fictional transactions, prepares a client’s chart of accounts in the specific JSON‑like format required by Jeb, and obtains an API key from Typesafe. He walks through the Jeb playground, showing how to feed a single transaction and a list of possible categories, and observes the model’s confidence scores. He then builds a simple Replit app that reads a CSV of transactions, sends each row to Jeb, and appends the chosen chart‑of‑accounts category and confidence score. Running the batch on 10,000 rows, he times the process, notes throttling and rate‑limit pauses, and finds that about 60 % of the classifications fall below a 60 % confidence threshold, with many obvious items (e.g., Microsoft) receiving low confidence. The total cost of the experiment is $0.13. He concludes that while Jeb is extremely cheap (≈$0.04 per million tokens) and fast per request, its accuracy is currently insufficient for production accounting workflows, and he would not recommend it without further improvements. He invites viewers to book a call for done‑for‑you AI automation services and points to a follow‑up video on building similar workflows with other tools.
Chapters
I'm about to categorize 10,000 transactions with AI, and the whole job is going to cost me less than a cup of coffee. Because the question a lot of accounting firm owners are asking me right now isn't just, can AI do this? They recognize how powerful it is. It's more like, how big is our AI bill going to get? And if you're running every single transaction through a premium AI model, it's going to get very expensive very fast, but not if you're using the right system. So in this video, I'm going to categorize 10,000 transactions live using a new super affordable model called Jet, and whether I'll build a production workflow for an accounting firm with a model like this. Now, the first step is getting the right files in front of the eye and the right data. Remember, I'm not sorting one month for one person. I want you to imagine we have 10,000 transactions, 5 years worth of bank transactions potentially. Of course, for this purpose, it's just example data, but that's the kind of work that would take a bookkeeper well over a week to categorize by hand and clean up and catch up. Now, to sort it with AI, we're going to need 2 things. The first is the transaction themselves, and the second is the client's chart of accounts, because every row has to land in an account that already exists in the client's books. Without that list, AI might just invent the category that sounds right, like, I don't know, office stuff, and now you've got a cleanup job instead of a finished file. So let me show you what it looks like to just get the transactions and the playground started to actually run this through this new AI. So what I've done here is I've asked AI to build me 10,000 fictional transactions for the purpose of this video, and I I can show you briefly what that looks like exactly, just to give you a sense of whether those are realistic based on perhaps typical transactions you would see in a bank like Whole Foods, insurance, electric, insurance, Shell, PayPal, local cafe, Seal, Microsoft. There's many popular vendors included here, and I can see here their full amount, their date, and this is dating from all the way to 2021, all the way to today. Briefly showcasing here as well that if I scroll all the way down, I can see that— well, it's going to take me a while. Let's just do it this way instead. It's, as we said, 10,000 transactions that we have available here for us to test with. The next thing I did is that effectively I selected a series of chart of accounts, a summarized chart of accounts, if you like, and put it on this specific format, which is the format in which the model will need it later on. I'm not going to get into the technical details of what this is exactly. This is the format, as you can see, with just 2 quotes, colon, another 2 quotes, and comma, in the order in which we want them to be in. So effectively, it will pick one of these categories based on the transactions above. I'm just going to copy that and keep it handy for now. All right, so now we need something to actually categorize these 10,000 transactions, right? Some sort of playground. And the obvious move is to throw them at whatever AI model your firm already pays for. But a premium model like Claude Opus or Fable answers by writing, and you pay for every word it reads and every word it writes on every single row, which means the bill grows with every client you add and every transaction, and nobody at your firm can tell what the bill is going to be next month. And that's where this new model, Jeb, comes in. Jeb is a model from a company called Typesafe, and it's built to do one thing pretty well. It's just to pick the right answer from a list. It cannot chat. It cannot write code. It cannot explain its answers. That kind of sounds like a downgrade, but that's exactly the point. That's why it's so cheap because it never needs to write anything. You only pay for what you send it and it's going to be like $0.04 per million tokens as opposed to like $10. So it's like 200 times more affordable. TypeState also says each answer comes back in under half a second and on every row it's going to give you a confidence score so you know exactly how sure it is about each category you pick. Let me actually show you how that works in the Jeb playground first. And then we're going to move to do this for all 10,000 transactions at once. All right, so where do we start? Let's go to— let's assume I don't know anything and just getting started here. I'm going to go to typesafe.ai and effectively going to go into the API concept. So if you're tapping this for the first time, you might have to provide your email address and confirm that. But effectively, there's 2 tabs that I care about, 3 really. Number one, it's the balance, and I already charged $5 here for us to test. But if you haven't done so, then feel free to add the funds at this point. And then I can go to API keys and make sure I create my first key. If I go ahead and click create key, let's just say this is YouTube testing number 2. I can create that key for myself and have it available. You should rotate it and not share this with anyone. This is a video, so for the purposes of it, I'm showing you how to actually get it. And here I'm just copying it. And again, just like I did with the chart of accounts, I'm just going to keep that handy somewhere. Okay, now I'm gonna go to the playground. This is where things get interesting because this is where I can at least start to experiment with the model and understand how it works. So I'll give you an example. This set of questions, there's different types of questions you can ask, and the one that's relevant for us today, it's choice, right? I want to give it a set of options to choose from. I'm gonna give it a sample text, which I can do it in plain text as you'll see, and it can give me back based on this level of confidence what's the right choice in this context. So what I'm going to do is I'm going to copy this. I'm going to put it on ChatGPT. Let's create a new chat so you can see what that looks like. I'm going to set it. Can you please fill this exact structure with the following chart of accounts as options? And I'm going to copy the list I had before. This could be technically just like a list, CSV, whatever format you have, effectively going to make sure that the format of the options is in the format that Jeb requires it to run. I'm just going to let it do its thing there, but that's the extent of what you should know about what the model needs to know what the options it has to pick from. Once we have it done, we'll just copy it and let's go back to Jeb, replace this section with that. And now we should be in a position to take the CSV we pulled up earlier and start just like dropping some transactions as examples. So if we have, as an example, let's just say Shell and I copy this, paste it here, I'm going to run. You see it's suggesting this Auto, which makes sense, right? It's like an automobile expense. If I do something like POS Office, let's see what we get. I'm going to run that. Office, again, makes sense. It has the word, so probably that's an easy one. Let's take Microsoft. What should Microsoft be? Software, I suppose. In this case, funny enough, it selected Sales and then Software. So again, you can see here has not necessarily perfect. Let's do a couple more. Home Depot, it's categorized as other or office. Perhaps we wouldn't know exactly how to categorize that one. Let's see if we can do Uber. In this case, travel. That makes sense to me. So you can see how regardless of the transaction I give it, it's giving me back the percentage of confidence it believes and the specific category it believes is the best fit for that specific description. Now, if you want workflows like this running across your firm without building them yourself, feel free to book a call using the link below. We build done-for-you AI and automations for accounting firms, and the way we work, it's simple. Every engagement starts with an efficiency audit that shows you where your team's hours are going before you commit to anything long-term. This way you actually know what AI can do for you, and that's how we've used the same process to find and save 3,900 hours of capacity for small and mid-sized So if you want to see what we can find in yours, you should book a call using the link below. All right, so this is where it gets a little bit AI inception, but I'm going to be creating a tool that would allow me to run this process through JEP. So specifically, I'm using Replit in this case just to build a simple tool that, as I say here, hey, provided a CSV and a chart of accounts— this was just a misspelling there because I'm using voice to type— in a format that the JEP model can process, help me build a simple app where Provided the CSV in 3 columns, the current columns they have, date, description, and amount. I want to add a 4th column of chart of accounts and another column of confidence that came from Jeb. I also want to be able to just input the API key to effectively have a tool that I can run all transactions through. And then I gave it the chart of accounts as I had it. It's kind of working at the moment, but we should be able to have an easy tool that even though this might cost me a dollar or 2 to create, now I'll be able to run this from Jeb and we'll see how it looks. Works. All right, so let's test this out. Supposedly, if I load the transaction, I should be able to see them here. I could test with 5 of them. You can see how quickly that went through, and their confidence score right here, which I should be able to download and then see in a simple format like this where I have the chart of accounts and the confidence score. This is looking pretty good, like exactly what I was hoping for. I'd like to time this actually, because part of the pitch of these models is not just affordability but speed. How quickly can it go through 10,000 transactions is the question I would have. So let's actually go ahead and time it, uh, here. I'll just do it on my phone. I'll just go tap at the same time, classify all 10,000. Interestingly enough, it's not going as quickly as I expected. I suspect that is because of some throttling at the code level. It supposedly can run, you know, infinitely fast, I've been told. Regardless, you can start to see here how sometimes it suggests very low confidence in certain areas, high confidence on others, like things like AT&T, water utility, etc. I'll just let it run and finish. All right, well, I am 9 minutes in and I'm afraid to admit that this has been a failed experiment. So it's still halfway through. It kind of rate limited many times, and probably after every 1,000 transactions or so, I had to get it to run again. And as you can see, out of roughly 5,000 transactions, so it was, let's just say half of them, probably 60% of them are coming in below 60% confident. I could download the partial result to at least look at what are some of those transactions. As an example, insurance, Whole Foods. Honestly, those transactions, I feel like could have done pretty well, like software, 44%, Microsoft. Like, come on, that's pretty obvious it's software. So I don't think I like the outcome of this experiment. And at least that is my take of the current state of it, at least based on this experiment. So would I use it within an accounting from workflow? Not right now. It's definitely a cheap model and affordable. And actually, we can just quickly check how many credits or how many dollars we spend through that experiment? And the answer is 13 cents. Makes sense. But the accuracy really wasn't there for me to be comfortable, at least not off the bat. It's possible that with more context and that with a better setup, it may perform better than that. I'd say it was a good experiment but not a model I would recommend at this point. So categorization is one of those things MiniAI can do for you, not Jeb apparently. Yet many firms avoid using tools like Claude because they feel too complicated. So if you want to see a simple breakdown of all the information you need to start using things like Squad Scales, Squad Projects, and Artifacts yourself, feel free to watch this video next. I'll see you there, and until then, stay efficient.
