Elkar AI Put to the Test in Live Financial Modeling with Honest Results for Modellers - Ian & Giles

Foreign.

The Mod Squad, featuring Ian Schnorr, executive director of Financial Modeling Institute. Giles Mail, humble MVP and co founder of Full Stack Modeler, and Paul Barnhurst, the FP and a guy.

We're going to be testing a tool called LCAR today.

The formula that it has derived is, I would say, the perfect simplest formula, one that is transferable to a person, that is dynamic and modular.

I think this is quite nice. The only one we haven't really seen give this much feedback. That was probably the quickest I've seen.

There are some white values, there are some hidden white values. There are some inconsistent formulas here. Right. There are some hard coded values and I really like it says I'll check. So this is a very confident AI tool, which I love.

That's pretty close to AGI, isn't it?

If you ask the tool the same question multiple times, you might get different response. There are some real meaningful errors and I think it's important to see what these errors are and how it did not understand what a model means.

Welcome to another episode of the Mod Squad, a series on AI tools where we test them so you don't have to featuring in Schnorr, Giles, Mel and yours truly, Paul Barnhurst. This week I want to let you know that I will not be on the episode. Unfortunately, I was not able to make it and so Giles and in were kind enough to record an episode without me on Elkhar. So we'll go ahead and roll that here in a minute, but I just wanted to let you know. So you weren't surprised why I'm not on my own show. I think it's great that I have two co hosts that can cover for me to super excited for you to get to listen to this episode. I think it'll be a real treat to hear voices other than my own. So we're going to go ahead and roll that. But again this week you won't have to listen to me, so enjoy it. I will be back next week. Thank you for being a loyal listener. And please, if you have a tool, we should test an idea for an episode. You want to come on the show? Reach out to me. I'd love to hear from you, the audience. You can reach me at P P barnhurst@the fpnaguy.com all right, we're gonna go ahead and roll this episode of the Mod Squad. Thanks again.

Bye.

Hello and welcome to another episode of the Mod Squad with me, Giles Mail. And I'm joined by just one special guest this time, Ian. Ian, where's our third musketeer.

Ah. He had an appointment where I was not at the last one. And Paul can't join us today, so we're gonna take over his hosting duties. It's not gonna be. I bet it was more joyful without me, but it will be less. Less joyful without Paul, that's for sure. So we'll do our best to make him proud.

We'll do our best. I mean, he's actually left us in control of a recording that will be going on his YouTube channel, his podcast channel. He knows how much I like to swear on recordings.

Possibly one of the least intelligent, riskiest things he's done in his career, but.

All right, Paul, you can trust us. You can trust us. All right, so I think we're going to be testing a tool called El Car today. Before we do that, I'll act as host for you. Where's your head at? We've obviously gone through quite a few rounds now. We've looked at Agent, we've looked at a fair few tools. What's your kind of general sense of where we're at in this series?

Well, I mean, it feels like we've. I hate to use the word plateau, but it feels like we're hitting a bit of a consistency. Right. We're continuously impressed with what the tools are doing. We'll see what if today's any different. I mean, overall, their performance has been. Would you agree, like on some of the tests you've been putting at it, more similar than not. There are various degrees of able to get some of the questions done that you're providing into it. And the e challenges in modeling, in building a full model, they have not been great to date at building from scratch, would you say? And we'll test this one. So I feel like we may have hit sort of a common consistent spot where we're at a bit of a hump and see if we can get over that. What do you think?

Yeah, probably the same. I mean, obviously you missed the excitement of testing Shortcut in the last episode, which was interesting because I agree. I think it's there or thereabouts. It's probably towards the better end of the third party. Tools didn't live up to its crazy claims, but, you know, I think Tab AI has probably been one of the most impressive. Did some really great things, but wasn't perfect. Agent has done the best so far from a sort of financial modeling perspective, I think. But again, I don't think any of us would say perfect or ready to rely upon, you know, in completeness. Yeah, I agree. I think coming into this episode, I was a bit like, oh, you know, I feel like we kind of know what we're going to get there or thereabouts. So I wonder whether that's going to change.

And, you know, I've been traveling. I've been on the road for the last couple of weeks, which is why I missed you in the last time. And talking to a lot of modelers traveling throughout the US And. And in London last week, and there seems to be a consistent feeling that it's aligned with our own views that nobody's ready to have a tool, build a model from scratch, have it done, package it up, send it in. People look are looking forward to using it as a collaborator, as a partner to help find errors, to help with the design process, to help improve models and get ideas. I don't think anyone's expecting or looking for a button, a magic button to push to get your model and saves you all that time and energy. Everyone wants to understand what they're doing. So that is what I would say.

I am hearing more and more stories of people and teams leveraging Copilot, mainly Copilot. I'm hearing and getting some value from it. So certainly not getting it to build models, but getting it to help look at data or to answer a query about, oh, how do I do this in M code, how do I do this in vba? So I'm hearing more of those stories, and that's a testament to the progress of Copilot, I guess.

Yeah. And I think we're hoping that in the modeling world we'll be able to say, use these tools to. To give me a suggestion and is there a better way I can think about modeling revenue? And do you have any ideas on what I can do with this depreciation calculation? And can you help me out with an idea? Right. I think we would love to see the same on the modeling side as well.

All right, good. I think we're still aligned, which is good. So should we have a look at Elcar? We'll bring up the website first and we'll just have a scan through. So this is. This is Elkar's website. I'll hand it over to you again. Ian, Anything that sort of jumps out at you as I scroll down here?

No, I mean, you know, again, I'm feeling like. I'm feeling like there's a consistency in the websites. I don't know if you would agree. Different colors, but they all tend to have the same A similar look and feel. I'm looking at it now with some, you know, animations to them, some screenshots. They're clean, they're simple. So I like them. You think it Elkhar. They all have a tagline and a, and a logo. There is what you think at Elkhart builds it. Their pricing plans seem to be in. Would you, you want to talk to. In line with what we're seeing?

I think so. So 40 40. And again, like I'm, I'm just not clued in enough to know when it says $40 and $250 requests a month, like exactly what does that mean? Is that 250 prompts? It might be, it might be something slightly different. If you had you planning to use this heavily, maybe you'd start to, you know, step into the $200 a month territory. And once you're there, it is a lot, isn't it, when you, you know, you can get Claude Pro or, or ChatGPT Pro subscriptions for 20, $30. So it is a lot of money.

What does it say there? If you go down? It says under the pricing for Max. So I have a feeling you're right. I mean you could, you could probably fly through 250 requests oh pretty quickly. I mean it's not. If you're working intensely, if you're working intensely in a spreadsheet, it's not hard to imagine going through, you know, 10, 20. If that truly means each individual prompt is a request, then you could go through 20, 30, 40, 50 a day. I would think. Pretty. Yeah, you're working on something iterate like iterating through it. So the max plan says 50. So everything unlimited requests, but 50 monthly requests in agent max mode. So I don't know what that means either. Does that mean that the best tool, Agent, which is powered by Opus, in the best tool, you still only get 50 requests per month? I'm not sure.

Yeah, I don't know. And it's interesting, isn't it? Because this gets to the conversation that comes up a lot in sort of MVP circles is like, you know, what about the cost of all of this prompting And I guess there is a significant cost behind the scenes for these companies as well. So I assume you're right. It's probably to link to the most powerful tools in the back end and you have to limit that.

So you're saying you're talking about the cost incurred by these third party add ins because of course as you put in a request, they're submitting a request to you Know, some sort of the LLM.

I think that would be my guess as to how it all works. So Yeah, I guess $200 a month, you could probably start to eat through that if you're looking at 50 requests in your kind of high powered mode a month. I don't really know. This is the interesting thing, isn't it? We don't know.

Okay, anything else looks good. Let's see how, how it adds. And you want to look at the features. Right?

I mean it, what I did, this is quite interesting. So they've got a bunch of videos here. They do actually have one for the Microsoft Excel World Championship. So this is going to take me to their YouTube. They obviously have an example of the Microsoft Excel World Championship case of some sort being sold.

Yeah.

So there you go. I'm not the first person to try this with Elkhart. They've done it themselves.

Yeah, absolutely. Well, let's see, let's see how it performs.

So I think let's get stuck in then. We don't have Paul, so when we're probably not going to do his test cases, we'll do mine and yours and we'll see how he gets on. So I will start with.

Great.

Start with the same one I have started with for every tool, which is the easier XL Esports Battle case. So let me bring it up on screen, just have a look at what we get in the LCAR kind of, you know, user interface and then we'll get going from there. So it is another kind of additional button on my home ribbon. I had a whole stack of these. I had to start deleting them because I had about seven third party tools in there.

Yes, as do I. So you just started removing them.

All right, so Elkar and, and for anyone watching, this is the first episode. This is an XL Esports case. You have some bonus questions and then you have seven levels typically where you'll, you'll given a problem to solve and then you have to fill in the answers in these green cells so that they get harder through the levels. This one is saying essentially pick the number of characters from the left from this word. So if you look at the example, three from Filtro would be fil. So we're looking for formulas to be created by the tool, like left with the number of characters we want it to link to these cells, ideally. And it gets harder and harder. So.

And ideally. And ideally. You know, one thing that we've learned in this process is I, I think, I would say ideally we wanted to build a solution that is one that could be transferred to a person to use and to copy. Right. And to automate. And what we've discovered with some of the other tools is that they are of course lightning fast but building tools that are incompatible with the way a person would work. Like they, they might. They might drop in a dozen hard coded values in a formula. Now it's okay. It can do that because it can do it instantly. But that doesn't leave it dynamic and. And modular and able to be copied. Right. I think we'd love to see it and we might need to change the prompts then to see. But let's see what it does. Let's just put the exact same prompts as before and see what it does hit again to.

So we just get it to do level one first just to see if it's got the right kind of sense of what we're doing. Uh, so we're already getting some feedback. I'll analyze the case tab. Perfect. I can see the structure. Level one is about. So again I think this is quite nice. The only one we haven't really seen give this much feedback. Oh wow. It's done. That was quick. Uh wow. That was probably the quickest I've seen any.

So you asked it. I'm sorry, you asked it just to do level one?

I asked it just to do level one. But it's. It has done level one. That was very well.

I love that. And so first of all what I like is it gave you. Can you scroll up for a second and show everyone. Go up above. It said where did it stop? Right. I need to create form. It's. I need to create formulas using left function and to like it's explaining there what it understands it needs to. That's great. Not all the tools were doing that, were they?

That's what I was. I was. I was saying before I got shocked by its speed. I think agent was probably the lowest in that kind of measure of. You couldn't really see what was what it was thinking. Yeah, continue. So let's do continue with the remaining levels.

And the formula that it has derived is I would say the perfect simplest formula also is transferable. Right. To a person that is dynamic and modular and it works. Let's see what it does for some of the more complex.

So I'll hit pause. Just because we don't know. It might take a minute, it might take four or five, but I'll hit pause.

Oh, look at that. Wow.

I won't hit pause.

You might not need. You just this is fast. Let's see here. Oh, wow.

Okay. Okay. So.

So. But again, I'm. Now help me out though. It's flying through. Which question is it on? That's impressive. The speed.

Okay, it's finished. Wow.

I know. We haven't been timing. That's fast, right? That was fast. That was.

I didn't have time to hit pause.

Okay, you did.

So there's a few things we obviously have to check. Annoyingly, I should have fixed this age ago, but I've got one where I don't have the green ticks available. So let me just pull up the answers on a different screen and I will just do a quick sense check to make sure it's got the right answers because obviously that would be nice. Yes. So right answers looking good. Looking at so far. Yeah.

And let's see. I'm looking as you evaluate them to see the caliber on the quality answer. So again, this one, I'm looking to see if there's 100%.

I have to say that is by far and away the fastest we've seen. Now if it was just about speed, it's one.

Let's have a look at what it did. Okay. So you know, you're familiar with these. It's put in some dead numbers. Is that fine? I'm not sure exactly. It's repeating.

I genuinely have absolutely no idea why. It's turned to the rept function and it's repeating that a hundred times. Right.

It's repeating this. And it is. Huh.

And it's. I don't know what it's doing, bro.

Look what it's done, though. It's. So this is what we're. This is not likely. It's certainly not likely. The simplest way to solve it and it's using. This is what I, you know, also keep talking about is I think that people's skills are going to have to increase just to understand what it's doing. This is a very unconventional. Would you agree? It's using an indirect. If you look where your cursor is, it's using an indirect to populate a row function. Now, a row function simply says, you know, Excel's row function simply says click on a cell reference and it tells you the number of the row number. What it's doing is it's using an indirect with a range from 1 to 10. You want to highlight that in quotation marks. But I'm not quite sure what that's doing in quotation marks. I'd have to test it. And you're the one Looking at it. And then it's embedding that into the row function.

So I have no idea what it's doing raviolist. I have no idea what that part is.

And you can see it's got a little array with those squiggly brackets, one, two, at the end. Would you see that at the very end of the formula? So that's all part of the sum product, right? It's all part. This would be. I think it's safe to say, I think it's 8%.

That's.

This would be a hard formula for most Excel users.

Okay. That would just. I don't know why it's. It's turning to that. But hey, it got the right answers in a really, really weird way level.

But again, so that. Here's the thing for those watching, you know that it got the right answer because you know what the right answer is, right? Just a cautionary warning. If you were asking it to solve a problem like this and you did not know what the right answer was, which is which? More simulates reality, right? In real life, you have a data problem, you say, can you solve it from it? You're not sure the right answer. If this is what you got, you're going to have to really kick up your Excel skills pretty significantly to be able to interpret and understand that, to make sure that it's actually doing it right and doing it right consistently. So that's pretty, pretty amazing.

Level three, fine. So it's used indirect and column. Yeah, that's fine. Okay, so it's done the job. Okay, there's level four again. Good. Len minus len substitute. It's using the cell references, which is great. I made one or two.

And that feels like the way. That feels like the way most people would solve it, you know, in a challenge. Right, that makes sense. Good luck with that one, Giles.

I tell you, it's so interesting that different tools that have sort of preferred approaches yet again it has gone for rept and blank and a hundred. I can't get my head around why it's adding a hundred times, repeating it a hundred times.

Like why. Yeah, I don't know why it's some. Maybe because that is again, significantly longer than what the total numeric chain is within any of them. I'm not sure. Again, I'd have to kind of go back and dissect it.

You actually, you might have a point there. Because if you did do mid and then if that was the second part of the middle. Do you know? I don't know.

I just Love if you start from the beginning. I just love. You know, for those who really want to elevate their Excel, you've got a substitute embedded within the mid, embedded within a trim, embedded within a value, embedded within a sum product.

But then just a bunch of hard coded numbers as well. 100 minus 99. I genuinely have no idea what that's doing.

So for those of you who've ever had your manager say, what the heck is that formula doing? Be prepared. Be prepared. Yeah, yeah.

So level six, good. Indirect on the map and the reference there, that's. That's good. And then level seven, if it actually looks more reasonable, that's that function again.

Oh, yeah, that. That seems that again, for anyone, for any good Excel user, that's not going to be hard to. To figure that out when it's. It's not a.

Okay, so I think that that is. I mean, that was really impressive.

Well, we didn't have to push pause. We didn't have to push pause.

We didn't.

And I think it's safe to say that. I think it's safe to say, right. I mean, in a comp. People would have 30 minutes. Is that right? To solve all of those. And most people would not get ready. Sorry.

Well, for that easier case. So I think, you know, that that's a case where the top pros would be finishing fast. They'd be doing that in, you know, 10 minutes, maybe less. But that did it in what, two minutes?

Maybe less. Maybe less. Maybe one. But. And that's. The very top pros that have practiced would get it in. Tim is. There are plenty of people that would not finish that in half an hour. I'm very confident right in of that one.

Yeah, I think so. If you. Because some people just wouldn't know the functions to use. So let's. Let's really test this tool. Let's go to my case.

Well, let's. Let's buckle up here because it did the first one in a minute. So let's see. Let's see how it does on.

If it does this as well, I have to say.

And then we'll test it on whether it can actually build a model. But let's see what it does on this one. This is.

Okay, tell you what, for this one, because I've asked it to do all seven levels, I'm going to hit pause. I can't imagine it's going to do it in 30 seconds. All right, we're back. We. We. This is probably the funniest thing. I think we've seen any tool do. So as we were going through, I noticed that it was putting the answer answers in the column here. And in these cases, this is where you get this really nice green tick. You see on level two. It actually put the answers in this column and overwrote the green ticks. So I moved it over and I also unhid the answers sheet, which has the answers but you hide it in the background. And I thought this was doing really well. And then when you look at the later levels, it's spotted that there's a sheet called Answers and it's just very smart to the answer.

This is the most clever AI tool.

I mean, we've looked at today. That's pretty close to AGI, isn't it? That's exactly what a naughty human would do.

We didn't realize you had an answer sheet. And it just said, heck, I'm not going to hurt myself. Let me just link directly to the answer.

That is absolutely.

Does it make sense? Does it make sense to pause again and go back and say, hey, can you, can you solve it yourself without checking the answer sheet? I mean, you, you could, but from the ones that it did solve, it looks like it's done quite well. Is that right?

I mean, I think it cheated for most of them.

Oh, okay.

So I'm telling you, I'm saying you.

Can'T use the answer sheets like scolding a small child.

Right? Absolutely right. I apologize for taking that shortcut.

See, AI can be fun joking around with the, with the AI tool. Yeah. Okay.

We should repause. We'll give it another go if it pause. It's just. It's. We're expelling it. It's done.

Absolutely.

All right, we'll be back in a minute. Okay, we are back. That was, that was a good 15, 16 minutes. It struggled a bit. So it's. Again, Ian, I find it really interesting because this one. So we have the little cheating episode, which we've forgiven it for, but just seem to be struggling with putting things in the right place. So it kind of got there. What did it do? It got level two correct. That's fine. Level three, a couple correct, but. But again, not, not exactly a full set of green ticks. Level four, it eventually got correct apart from one, and actually that looks correct in terms of some of these kind of characteristics for the coats were doubled. Some of them were zeroed out. So up until that point, I think it did well. Level five, it just put the same calculation across all the levels, which was wrong. Level six Again, that's wrong. And then again, I'm not giving it this because I think it's still trying to cheat. It's just hard coded almost the right answers.

You're looking, you're looking for it to. Yeah, like again, it's. It's doing something that it can do as a computer, but it's not a sort of called a human solution. Is that what you mean?

Well, I also mean that I'm not sure I totally trust it. After our little pulling the answers from the answer she escapade, I think it might just be looking at the answers and hard coding them in. In a formula. So I don't know. So this is a harder case. We knew it. I almost got the sense if this was a human analyst, after the speed of the first case, I almost feel like I would be saying to this human, just slow down a bit. Just, just take your time. You don't need to rush this much. Like, it's more important that you get the right solution and you think through the logic than you win this race.

Well, the other thing too is we all know, we all know from any Excel work we've ever done that Excel problems and challenges tend to elevate in complexity, which is what's realistic about the esports challenges. I mean, in life, often the first things you have to do in any spreadsheet are just add up a column of numbers and then, you know, and then do some additional more complex calculations and eventually build in lookups and other. I mean, it just, it elevates. And what we're seeing is these tools are good at doing the very easy and then medium complexity. And as you get into more and more complexity, they hit a bit of a wall. And this one has as well. It appears right?

I think so. But it also feels like it's just like it's had a little malfunction in ways that are quite unique. You know, we said put the answers in column E, it stuck them all in column F. And we've seen bits like that on a few tools. There's a very similar characteristic where, you know, one. I remember one of the tools for the financial modeling challenge that you set. It tried to do the balance sheet check, but one of the line items was pointing to the row above the total assets calculation. You're like, okay, that's just linking to the wrong thing. So, yeah, I think it's done okay in a very, very hard challenge. Uh, should we do my other one? Should we do my other one quickly and then we'll move on to yours?

Yeah. What was your other one?

My other one is the trial balance. So I'll tell you what, why don't we come back to that? Should we just to. I get bored of my own voice. You wouldn't believe it. All right, well then let's skip over to you.

Sure. Why don't we do that? Should we warm it up with some. With doing some checks and some testing and looking at a model again? Why don't we do that? And then I'm gonna. I really want to make sure we have time to see what it will do. I have not, you know, full disclosure, I have not asked this tool to build anything from scratch yet. So this is going to be a real time surprise for both of us when we see what it can do. So let's go here first of all, and I assume that you've got my screen up now, is that correct? So you do. And let's start. Let's start with. So here is the LCAR tool. I'll get. Load that one up. You can see I've got lots of them here. And it. Admittedly it took a little bit of time to get this one working. Sometimes tools take a little bit of effort to get them loaded up, doesn't it? This one, some people might not have had the. The patience, but, but we did get it working right. And it, it is working now. So it confirmed that it does work. I'm gonna. It's asking me to upgrade. I have a week still. I'll just. Just to get rid of this. So let's try some again. I have a financial model here and it's built as you know, the balance sheet is not balanced. The balance sheet is out of balance on purpose and it did not watch me. So I'm curious to see how it solves that. There are some white values. There are some hidden white values. There are some inconsistent formulas here. Right. There are some hard coded values. There's some. Must have. Let's see, there are some hidden sheets. Let's start to see number one. Are there any hidden sheets in this file? I think it's going to solve these fairly quickly based on the performance. And I really like. It says, I'll check and it says, wow, look at that.

It's quick.

Yes, there are two. So this is a very confident AI tool, which I love. Right. It works fast and when it gets the answer, it says, yes, it's got it. So there are two hidden sheets. Covert and clandestine. Done. Easy. What about this? Let's see if there are any. Let's ask it if there are any white values on this sheet. And it says are there any white values? I'll check for text values on the model sheet by examining the formatting. So it understands that it's going to check the formatting of all cells and let says. Let me try a different approach. It's doing something. I'm not sure what it's doing. It's trying a few times. It's looking for. Okay, well, look at this. This is interesting, Charles. It says no white text found on the model sheet. The font color check did not detect any cells with white text in the sampled range, which was B1 to U50. Now U50 goes way out here. However, check it out.

Yeah, interesting. So just go that wrong.

I can see there's some numbers here. These cells were white. And again, I'm not sure why. There are absolutely white cells here. You can see they're white. So it missed that. I'm not again, not sure exactly what it was doing it. But obviously needed to be careful about that. Not. But it clearly understood what I asked. Right.

I think that's what's even more interesting. It very much understood the question and then still didn't spot what was there. Which is.

Yeah, well, what about this? Are there any inconsistent formulas on the sheet? That's a key one. Are there any inconsistent formulas? And I know, I'll check and I like here, I'll check for inconsistent formulas on the model sheet by examining formulas across similar rows and columns. I like that. It's scanning the model for inconsistent formulas. And it says yes, this is great. This is fast. This is great. Yes. There are 13 columns with inconsistencies. The main issues are column B not as worried about column B because column B is sort of just a bit of a helper column. A text column. Not as worried about that. Column C to O. It recognizes some things. Now that's interesting. And again, this is something we're finding as well. Right. And I forget how people refer to this issue. I'm drawing a blank right now. I should have the word. But the issue being that if you ask the tool the same question multiple times times you might get different responses. Because I tried this one earlier and it did. Earlier it did identify that row 13 as an example was a row that had an inconsistent formula. This one is different than the prior one. It knew that. But this time I asked it the exact same question as I did, you know, an hour ago. And are there any inconsistent formulas on the sheet? It's saying yes, last time it Actually showed me the rows that were problematic. This time it's not. So that makes me uncomfortable generally with AI tools as well that we've seen is I always wonder, okay, it's got it. But we'll always get that. And how many times do I have to try running it to let me. Let's try again. Let's just literally run it again. Are there any inconsistent form? I'm literally deliberately not changing it because I know that that same prompt gave me a different request last time. So, you know, does it learn from itself? Is it going to do the exact same thing thing? It's checking for actual errors. Wow. Now it says suspicious patterns detected. The formulas on this sheet appear to be valid, even though there were multiple formula patterns. So this is concerning. So not only did it still not pick up the answer, but now it's telling me that there are no formula errors or suspicious patterns. And we know that there are. I'm going to do it again. I'm going to try it one more time.

Interesting.

There is that I'm not telling it that I think it made a mistake. I will do that next time. How about that? The next time I'll say, I think you're missing something. Because I just want to see what happens if it runs it over and over again with the exact same prompt now, because if you ran it once, if you ran it for most people, I think, Right, Charles, if you ran it once and you got this, you'd be like, great, that's awesome. Right? No inconsistent formulas. And then your boss would discover it at midnight.

Well, exactly. And that's. This is where my head's at with my kind of worries at the moment with this stuff, is if I used a tool like Oak, which I've used for 15 years, I'd pull up the model map from Oak. I'd be able to scan the inconsistencies visually in about five minutes myself across a huge model. And I'd be right.

Yeah. I mean, let's be clear. You don't need AI to do any of these model checks. But it would be nice to know that it could do it for you here. Right now, for the third time, it's given me a third different response. Yes. There are eight columns with significant formula inconsistencies. Again, strange, because just one minute ago, it said that there were no formulas or errors or suspicious patterns detected. Right. And, and, and now it's looking at.

It as a wrong axis. It's just. It's going vertically and saying, oh, within a column, that there's differences. Which is.

I'm. Now let me try and challenge it. I'm going to go off script here. Are you sure row 13 has a cell with hard coded values? Are there other rows with formulas like that? Okay. I think they don't like to be challenged.

Are you sure?

But let's see. I'm asking it if. If it is sure. It's. Are you sure? Row 13. We know row 13 has a cell with a hard coded value in it. Are there any other rows with formula? Let me check for hard coded values in formulas across the sheet. So it's trying. Yeah. Okay. It's very proud. It's very proud of itself. It's found it in M13. Well, I did give it a hint. That's true. That's true With a hard coded value. Let me search for others. Okay. So it is. Obviously I'm being a little bit difficult with it, but this is the kind of thing I would hope. One I don't remember off the top of my head. If there are others. Well, look here. Right here in O13 is another value with a number five at the end. I don't think it's pointed out that O13 did it find Oath. I don't think it did. Right. No. So I'm not sure how well I'm going to sleep at night if I know that this is the caliber of the checking for hardcoder values. But you know, anyway, it's. It was good to see that. Why don't we try moving on to the next one? How are you feeling about this one, Giles? I'm going to ask it why the balance sheet is not balancing. Are you feeling comfortable about this one? Are you a little more nervous about this one Mix.

And again, I feel like you talked about this plateau that we've reached. It's kind of aligned to what we're seeing. I think it's done some good things. It's done some really bad things. And that's like.

This is a hard question. This. Exactly. And I don't know. Listen. There are not many modelers out there in the grand scheme of things who could look at an unbalanced balance sheet and very, very quickly instantly determine you got to be. You have to think through a lot of things and understand a lot of system integrations to understand where to look and why a balance sheet might be out of balance. You have to understand a lot about accounting. So. But let's see. Let's try. Why doesn't the balance sheet balance? Let's go populate that one. It's a simple ask. It's a simple ask for a difficult problem. Right? Because of course, there could be 20 or 30 reasons like not only. Right. And again, the balance sheet's out of balance. And just as a reminder, so I fully acknowledge this is hard because my error is in the working capital schedule. And by the way, the change in working capital here, it's not that it's wrong. The change in working capital like this actually is the change in working capital. The change in working capital went from 56 down to 49. And so it dropped. The change in working Capital dropped from 56 to 49, it went down 6.9. So technically that's not the wrong answer. However, any, any financial modeler knows that when you're talking about change in working capital, the way a set of accounting statements prepares it, this is referring to the change in cash and as a result of the working capital changes. So for a financial model, we all need to know that you have to change the sign. I don't care about the actual change in the working capital. I care about the cash implication as a result of that change, which means that I have to subtract it the other way. Last year, minus this year. It's got to be positive. Now I'm just talking as it works through.

Let's go through it. Yeah, I think it's done.

So the balance sheet doesn't balance sheet. I'm just seeing the balance sheet doesn't balance because of discrepancies in the projected years. Look. Okay, thank you. So it knows in row 188. It realizes here that looking at the projected years in row 188, it knows balance. Okay.

So it is identified the root cause.

Well, it says the root cause is the hard coded value in M13 and M14. The hard coded value artificially inflates the net sales price, which cascades is using big words like cascades. Right. And it's. Here's, it's. It sounds intelligent, very intelligent. Right. But the reality is any good modeler knows that you can hard code revenues till the cows come home. That has nothing to do with an unbalanced balance sheet. It's right, it would just if the model was working properly, it would lead to inflated revenues and inflated inflated revenues, which would mean higher cash and higher retained earnings on the balance sheet. Right. That should have no impact on an unbalanced balance.

I also, I just think it doesn't know and I think it's pointed to the thing that you referenced earlier because you mentioned specifically that there was a hardcoded number, 50 something. I think it's just almost just gone there because it's the most logical guess to go to.

Because, I mean, let's just prove to. Of course, I mean, again, a modeler understands that, you know, you can. There are many mistakes and models that lead to an unbalanced balance sheet. There are even more mistakes and models that have nothing to do with the balance sheet and will not lead to a balance sheet being out of balance. Right. I mean, I'll copy this across. This is correct. I'll control R and copy this over. And my revenues, you know, we're still going to stay, you know, significantly out of balance. And just to prove I can double, I can triple the revenues. If the balance sheet's working properly. It will just lead to higher. Because of course that makes no sense if you double count revenues and it. You're going to have higher revenues, higher net income, higher cash and higher retained earnings. As long as it's built properly, the balance sheets stay in balance.

Yes, I think it. I think it's got that wrong. I don't think it knows. I appreciate it.

Are you sure? Let me see. Are you sure? Don't you think maybe. Let's see if we can. Don't you think the real error might have something to do with the working capital calculations? Are you sure? I like. You're right. I love it. You see that? It says you're right again.

Like to me, if I was to put this in like a human avatar, I would. I just imagine this, like really hypertenic, like, you know, late teens, early 20s. Person is really clever. It's like just slow down, slow down.

Well, look at that now. Now it's massaging my ego here. It tells me I'm right and then it tells me. Excellent catch. Thank you. But hang on. Aren't you supposed to be the one doing the catching, the IT tool? Right. I feel like my AI tool is supposed to be doing the catching.

It's got it wrong.

But anyway, I. Yeah, now it tells me I found. But it does think it found the real issue. The problem is in the working capital calculation. Look at the formulas. K260 to K2 65.

Use.

So it says accounts receivable use cost of goods sold, which is fine. Which is normal convention.

But surely if it's got. Accounts receivable would be linked to revenue, wouldn't it?

And accounts the Incans. And then it's. Here's a. This, this inconsistency creates a mismatch. The historical years, days are calculated as about Right. Well, it's sending me down, it's sending me down a, a bit of a. An incorrect path here. It's sending me down chasing something that's not right. I mean this is all fine. The days are calculated as the historical and this is a correct formula for days and the projected years. Balances. It is normal to calculate. Let me show you. Okay, so it says that the working capital balances should be calculated using revenue, not cost a good sold to maintain consistency with how the historical balances. This is causing the change in working capital to be incorrect. I really am allergic. I'm very, very hesitant. And it causes me a lot of. I don't know, this causes me a lot of. Lot of unhappiness when it over confidently tells me that yes, it found an answer which is not right. Right. So this overconfidence when you're wrong. So it's pointing out that the history. Look, standard convention dictates that most people use the standard convention of calculating receivables off of revenues. And you can see here in the future we've calculated the receivable balance by using net revenue because accounts receivable are most closely aligned to revenue and everything else is typically related to cost of goods sold. That is standard industry convention. You'll notice that historical days also we use cost of goods sold. It doesn't need to be. This is just standard convention. You don't need to. But. But also it has incorrectly said that this inconsistency is what causes a mismatch. That's not true. Right. I could calculate the historical days off of cost of goods sold instead of off of revenue and I'd get different numbers. It's not that, that's not standard convention. But it would work. It wouldn't matter.

It would accurate but it would still work and give you a result and not create an exactly like.

Let's prove it. Right, let's prove it. And again I will move on. But again it is really worrisome that it takes people. It will take people without knowledge down. So let's just show again the change in working capital. I have current minus the past year. That is the change in working capital. But what I need is the, the, the reflection on cash. So I either have to flip the sign here on the change in working capital or, or I can keep it that way and on the cash flow statement. Flip it. My preference would be to flip it right here. So I am going to take last year minus the current year and that should.

Oops.

That should get me to flip the balances that now the working capital has gone down, but as, but the cash has gone up because it means that we've sold off some working capital and put that cash into our bank. And of course now the balance sheet of course should be balanced. But it makes me very uncomfortable that on two occasions it suggested, you know, a course of action that's just not correct. And a modeler has to understand that or it's going to be a lot of painful nights. Let's try this. Let's go now and use up some of our additional time. I'm curious to see what it does. This one we might need to pause. Right, we might need to do a pause. Let's try this one here and let's go back into this tool.

So this is the one you're going to ask it to build a full on forecast, right? Exactly.

I am just going to say let's see what it does again. Build a five year forecast model on the model sheet for the years 2025 to 2029. Build separate schedules for revenues, costs, depreciation, income tax, working capital, debt and equity. Make reasonable assumptions. And I think that's a fair ask, you know, just to make some reasonable assumptions. We saw some tools did. Did okay, and, and most are not doing that. Great. There's a lot to understand. But let's, let's just click enter. I do not expect it will be able to generate a response instantly. So let's just see what it says and then take a pause. I'll help you build a comprehensive. And then what I'd love to do because I think we have a few minutes is we'll come back to the other one and we'll see if it can do evaluation. This for the first time. Let's try and get a bit of evaluation going. But interestingly, look, look at this though. Do you see what's happening? Maybe we should actually keep this going because unlike the other ones, it's actually building as it goes. Right? Like, look at this. It's doing things and formatting them. Look at that. It's building something and formatting the revenue is. Okay, okay, well, look what it's doing here. Wow. I mean this is impressive in speed, but again, this is the teenager that you want to scold. Right? Because it's. What it's doing for revenue is. And again, I said make reasonable assumptions. Totally fine, fine. So I'm okay with that. But I also asked it to build a separate schedule for revenues, cost and in this case it is just applied A. Well, it didn't. So it made an assumption for 5%, but it's not. You know, what it could have done very simply is just made one row above at the top that says, you know, growth, revenue growth, 5%. I mean, I would hope it knew that it's not ideal to type in a dead number like this right in. But okay, it did it. I mean, I guess I could modify that later. What did it do?

The numbers look broadly reasonable. I mean, albeit the tax, the interest expense has dropped a lot. It'd be interesting to see what it's done with the.

Well, look what it's in, though. Cost of goods sold, it's growing at 4%. So most modelers would understand and appreciate that there needs to be some connectivity between the cost of goods sold and the revenue.

Right.

In this case, it's literally just growing cost of goods. I mean, I sometimes joke when I used to teach a lot of modeling that anyone could build a model in literally five minutes. You can build a model of any company for five minutes. If you literally take the last year's numbers and apply growth rates to everything, then you've got a model. It doesn't mean that it's a good model or that it's actually reasonable. But in theory, you've got to forecast.

That's always.

That has been your bugbear about just.

With the hype, the big statements. If it can build a model in 10 minutes, it's like, yeah, but what's the model? Is it useful? Can anybody get value from it?

Right. Remember, we always talk about the fact that a good model has got to tell a story. It's got to get you smart so that I can come back to my boss and my client and my partners and say, this is a really smart thing. Forecast. Well, they're going to very quickly realize all I did is I grew revenue at 5%, I grew the cost at 4%, I grew SGA at 3%. We even depreciation is just growing at 2%. I mean, interest is growing at 7%. Oh, no, it is. Interest is. Sorry, I'm sorry. It's not growing interest, it's taking. Oh, look what it did, though. You're going to love this.

Right? No, no. So it's hard coding the principal repayments period by period.

It's a hard coding. No, it's hard coding the debt amount, I believe. But it's, it's.

Sorry. Yeah, that's what I mean, though.

It's like 175. It took a debt amount of 175 which it found. Well it's the end of the same year. It's taken this, it's got a balance sheet. We're going to quickly see if this balances jobs. We didn't even have to pause. This is impress so the speed is impressive. It's taken. It's taken. But instead of linking to the end of the balance sheet it typed in a dead number. It's applying 7%. Okay. I mean that's not crazy. And it's got an interest expense it what it do for taxes. Look at that. It built a little if statement if 21%. This is still very, very impressive and I'm sure I could modify it says if the net income if earnings is greater than zero then apply a tax rate of 21%. I, I bet I could ask it to, to just take out the hard values and put them in as assumptions. But we could try that the net income. So it's done this. Let's see here. It's good now let me continue. I wonder if it's done. It's, it's, it's been updating me as it goes. It's been the model. Okay. So first of all key assumptions. Oh then it claims it's put the assumptions in here. So this is really actually very cool. The net income, it's linked that up correctly. The depreciation is linked that up correctly. The deferred taxes and that's impressive because these are all historical numbers are dead change in working capital. It is taken by 11. Well I'm wondering what it's doing here down here. What did it do? J111 what is 1J well, J1. So this is. Oh, it took. Oh J11. I'm sorry J11 minus I11. It's taken the difference.

It's taken 2% of revenue is the, the change in.

Okay. As the change in working capital. I would still think though that logically if revenue is growing you'd expect that you need to invest more in working capital. So you'd expect that that would be an outflow right that that should be an outflow of cash because you're spending cash to keep up your, you're needing to, you're needing more working capital. You need more inventory, more receivables. Whereas this is a positive number which means that we're generating cash every year. Okay, whatever.

It's a weird way to go about it, but it gets to an answer, I guess.

Yeah. It's got a capex figure. It's taken capex. It's taken A revenue it's taken 6% of revenue as the capex assumption okay I'm not now this these. It's hard entered in some the long term debt it just typed in dead numbers I would have loved to see change in the long term debt from below but this is amazing that we actually get to audit this model on the fly right now let's see here.

Even the fact that it's got the structure looking right puts it in the top bracket of what we've seen because there's only been agent and I think shortcut did a reasonable job to get the structure right even if some of the logic wasn't there 100% so what.

Do you think Giles? The historical balance sheet balances and you let about the future Drum roll please.

I didn't think it has night oh Dials Giles.

Oh.

Oh. Oh dear.

What's going on here? Yeah I mean it's only off by $37 million. It's really not a big deal is it? Most people wouldn't care. Yeah listen so it is again I'd have to dissect this further it is off to a nice start. It is. What is it done though is it's told me the assumptions it has used so it has used assumptions. It says it put the assumptions in rows 112 to 137.

All right. Is it just here? Okay.

But as well it's funny. It's so interesting so what it did is it's. It put the assumptions in on an assumption but it's not using them anywhere. Right. Like you can see here it's not so it entered 5% but if I do a price dependent it's not going anywhere. Right. So it didn't. It didn't use that assumption let's try one more thing and let's get to do some valuation this is. Should I compliment it? Should I give it a compliment like it gave me one? This is great. However, can you remove the hard coded values from the formulas and link to the assumption cells instead in rows? What did it say? 112 to 137 so let's ask it to do that let's ask it to replace and remove the hard coded values that it put in and see if it can use these values I like that it actually has Right. I have no problem with this. I told it to make an assumption and it did for revenue growth, 5% freight and warehousing growth and it's actually described what it's doing like it's a. This is it knows that this is a growth rate. This is a growth rate. This, these are all growth rates. It's told me that the 21% is a tax rate. It's assuming that 10%. So again, this just not how deferred taxes would work. I get it. It made an assumption. But I would also hope that it would know how to make a reasonable assumption. Typically, you would not assume deferred taxes as a percentage of the cash tax and of the current tax. You know, so it has done something. Let's see what it's doing here now. It's doing a lot of work here. Let's see.

But again, I just can't get out of my head the fact that it's like, if this was a junior analyst, you'd, you'd be saying like, hey, look, you've shown some really strong signs that, you know, the good signs, you got a lot to work on. If it was a tool where I knew that the more time I invested, the more it learned and remembered, I'd be in, I'd be invested in this process. But I'm not even sure we're there yet with these tools. I don't know that they remember everything. So it's kind of dead time if it doesn't get things right.

I. Yeah. And you know what? While it's. While it's looking. While it's looking. So I don't know if. Let's see if it's doing that first of all. Let's see. Yeah.

Okay.

Look at that.

It is good.

So bang, it did it.

I'm kudos. Like, it's one prompt. Like, you literally just have to say, hey, link to the assumptions and it's done it. That's cool.

And by the way, I'm going to go into the formula just to double check it. I'm going to select the cell reference and press F5 Enter. It'll take me right to that cell reference. Yeah, it's good. It's linked it up correctly. So I'm happy with that. It's now. The model is now fully dynamic. Right. Let's see. Can you make the assumption values blue. Let's see if it can do that. Let's say I should have said blue font. Well, look at that. I didn't even have to. Fantastic. So now there are assumptions. That's model blue, which is perfect. So I wasn't even as clear as I would have liked, but it was. So this is good. Now, of course, we still have a situation where the balance sheet is not balanced. Balance sheet is in balance. Any ideas? So I'm going to try writing to it in a more, you know, slang. Colloquial. Colloquial way. The balance sheet is in balance. Any ideas? Why not? Let's see if it can follow that, you know, along. Obviously, prompt writing is important, but it seems to understand. Okay, let me check the balance sheet to see what's causing the imbalance. So I. I like that it understood my English.

I found the issue.

I found it. It found it. Found it. Quickly. I found it. The balance sheet is imbalanced because total assets. Oh, Giles. It's because the total assets doesn't equal the total liabilities.

Oh, this is my favorite tool.

That's why it doesn't balance, Giles. Because the assets don't match the liabilities. The problem is that the cash position is going very negative, which is true. Is creating a mismatch. So what I suspect is happening here, actually, and I didn't want to say. Okay, so it's actually made a bunch of.

Oh, it's still changing. Okay.

Ah, look what it's doing, though. I knew this might be the case. So negative. Revolver. Giles. Wow. Giles. Wow.

Oh, yeah. So it did fix it, but just. I just spotted the revolver line. So if this is the balance sheet, the revolver is negative.

Well, so. So impressive. So impressive. Now, first of all, it's done something. I haven't gone through it now. Look at that. It's getting the last laugh on us now. It's probably laughing to the bank at us. It's balanced. That's great. It has. I'd like to see what it did. It's got a. It's got a revolver. Now, it knows that it needs to have a revolver formula, which it did, but the revolver formula is negative. So that does. Oh. Oh, oh, oh, oh, oh.

Is it just plugged?

Oh, Giles. Oh, Giles. It plugged the balance sheet. Giles. Yeah.

This is a cheat. This is a cheese.

It took total assets minus everything else. It took the total assets minus everything else.

Can't have that Alcohol. That's not on.

No, Balanced it by. It balanced the balance sheet. And the funny thing is that it chose to do that on the revolver, which ended up making it negative because the assets are less than liabilities. Of course, now, I'm going to eyeball this to see. It does appear that the debt is going down. These numbers are reasonable to me. But if we look at the cash, change in cash is positive. So. So there shouldn't be a massive need for a revolver. Now, I'm not going to spend why don't we do this? Let's pause it. Let's pause it. And then I'm going to spend a minute trying to figure out why it's out of balance, and then we'll come back and show that and do something. Why don't we pause it now and.

I'll spend a minute. We're back. Ian, how's it done?

We're back. Well, this is. This might be the most impressive learning. And I think this is a really important part for anyone to watch to see what's going on here. And this causes me concerned about modeling and the dangers of two things. The dangers of having AI build a model without. There's a lot to understand. And number two, the dangers of a human not understanding what is going on. So Giles and I, we spent literally, was it eight or nine or maybe 10 minutes going through it and figured out all the real reasons, if you recall. I know we laughed. We had a fun laugh at the fact that what it did. So of course the balance sheet is not balanced. It's. Balance sheet's not balanced. And what it did is to balance it. It stuck in a plug. A classic area put in a plug. Basically, it took total assets minus everything else, and it ended up with a negative revolver. That's a disaster. There are some real meaningful errors, and I think it's important to see what these errors are and how it did not understand modeling, what a model means, and the integration of a company's financial statements. So I'm gonna do this very, very fast. Working exactly off what it gave us and where we got it balanced. And so we noticed that there were a bunch of issues. We found them very, very quickly. One of the things we noticed was this one of. So we first of all noticed that it had some links correctly, right? The. The. On the balance sheet, cash was linking to cash on the balance sheet. That's great. Receivables. It made an assumption that the receivables were just going to grow. Oh, sorry. The receivables was going to be a percentage of revenue. Okay. Not standard to do it this way, but not the end of the world. It took revenue and it simply said that revenue would be. It said that it would take revenue, and it used an assumption here of 12%. It said 12% of revenue would be outstanding as AR. I'm okay with that. And it did the same thing for inventory prepaids. That's how it arrived at the working capital. Now, if you look closely here, if you look closely, the receivables went down, the inventory went down, the prepaids went down a lot. When receipt, when assets go down, cash goes up. So just by these three going down, that should mean that there's a lot of cash coming in. Let's go check the change in working capital on the, on the cash flow statement. Well, to our unfortunate surprise, it didn't appreciate or recognize that there's a connection. It didn't understand that there needs to be connectivity between the balance sheet items and the cash flow statement. So it literally calculated the change in working capital as the difference in revenues multiplied by a percentage. It took the difference in revenue multiplied by 2%, but it's not connected. So what I did when we were offline is I, I now in real life, I would tell people never, ever, ever to do what I just did. Never do this in real life. You should build a working capital schedule line by line, nicely organized and link them in. In fact, I should just show it here. Right. This is the, this is the completed Henderson model, a real good model. Should have a nice working capital schedule. Oops. With all of the lines laid out very clean for receivables, inventory payables linked into the balance sheet and the change goes into the cash flow statement. Well, what I did here just to get it working is I did all of that work in one cell. So I just took the difference in the receivables, the difference in the inventory, the difference in the prepaids, the difference in the other lines, and I'm going to copy that in here. So really the change in working capital should have been 17 million in the first year, which. Right, that, that's what the answer should be because the assets went down a lot. So there should, based on, based on what it did to the assets, that should lead to a cash infusion. So that was one big mistake. It did not understand that there needed to be connectivity between them. So I put the proper formula in and I copied it over. Okay, fine. This was good. It also calculated for the CapEx. It's got CapEx. It's also taking revenues multiplied by 4%. Okay. To get at CapEx, it has another row here called other, which we're going to come back to. It's showing some other cash inflow of 2 million and then a negative 6.6 in investing activities. Fine. In the financing activities, it's got some, this is a formula, but this, whatever it's doing, it's simplistic. But let's see where the real problems are now on the balance sheet. Well, on the balance sheet what we learned was that the way it did PPE is it took last year's fixed assets and it minused the depreciation. Right. J20 is the depreciation. It did that correctly. But what any good modeler would know is that PPE has to be the previous year minus the depreciation. But add the CapEx, you also have to add the CapEx. And it didn't do that. It did not realize that every line on the cash flow statement has to hit the balance sheet somewhere. So let's pick that up. So that was a big one. So of course the PPE dropped by the depreciation but increased by the amount of the new spend, the capex. So again, worrisome that it knew, it knew to have a capex line but didn't realize the connectivity of where that needs to go on a balance sheet. And now since we saw this row as well, it's going to make me wonder. You know, one of the biggest all time mistakes people make on balance sheets is they don't realize that every line on the cash flow statement has to hit the balance sheet and the change in every line on the balance sheet has to hit the cash flow statement somewhere. So there's another row here called other. And let's see, it's not going anywhere. It's only going into the total. This 2 million is not going anywhere. It didn't pick that up. So let's go put it somewhere. Let's just stick it into ppe. If there's a cash infusion coming in in investing activities, it means an asset somewhere was sold off and they got some cash. So I have to reduce one of the asset lines. I'm going to just reduce PPE by that equivalent value on the cash flow statement. If it's, if cash is coming in, it means that there was an asset somewhere that was being decreased. Okay. And then I've got the other which is now picking up in my working capital. Let's just see what is now. Now instead of being off by 31, I'm off by 7. I know that this is another one. Another connectivity issue. Right. Giles, we saw the deferred taxes it took. It calculated the deferred tax liability as just the total assets. This is unusual. You'd never do this. It took total assets multiplied by 3%. That's not how you calculate deferred taxes on a balance sheet. On a balance sheet it should just be the prior year plus the deferred tax expense. The new deferred tax expense on the income statement. But you need to know that. A good modeler needs to know that. So I linked that in and copied that over. And now there is one more thing here.

Retained earnings. Yeah, I think there was something retained earnings, wasn't it?

Exactly. Remember what it did. It took the prior year, it added in net income, it added in net income, and it subtracted the dividend. In theory, correct. Remember that? But it subtracted a dividend. That's a negative number that this is correct. On a cash flow statement, the dividend would be negative. But you can't then subtract it in your retained earnings. By minusing this dividend, we're adding it because it minused a minused. So it made the retained earnings go up, when in fact the retained earnings needs to go down. It needed to the formula needed to add the net income, but add the dividend because it's already negative. And if we do that and click Enter now. I don't believe I. This is important. I do not believe that the AI would have been able to. Yet. Yet. Maybe in the future, maybe down the road. But, you know, I really believe that what I just did, a good modeler needs to be able to check audit, understand, and be able to really quickly identify what's going on and not simply rely on the AI. So I'm really pleased. Why don't we end it here then? Jazzle? I'm very pleased with what it did. I'm very pleased that it was able to get me a start. But as I've been saying, you know, it's got to be my partner on my shoulder that I can bounce things around, not send me down, you know, rabbit holes. And I would hope that it understands the system and the integration. But anyway, we're off. It is a good start, but it needed a lot of us to kind of get it working. Would you agree?

Yeah, I think from my side, to wrap up, you know, we spent what, 10 minutes, the two of us, very experienced at looking at three statement models. We found it, but we know what we're looking for. And again, my. I'm in the same place I've been for a long time with this, which is. I love where this is going. I think the future is definitely that these tools will make our lives so much easier. But do not listen to the hype saying, like, it builds financial models. It's gonna like. It's just. It is building something that looks like a financial model and you have to know how to troubleshoot it, otherwise you're gonna get in a world of trouble. It's probably one of the best.

So two things for me. Number one, right now, it's getting a good start. It's not good enough. The model that it built is what a brand new, you know, modeler would do. It made some classic modeling.

Absolutely.

Will it be able to fix those over time? Yeah, I believe it will. It'll be able to learn. But I also strongly believe that your boss, your client, you know, your colleagues are still going to say, giles, what did you do? How did you assume Capex and what did you do and how did you get the. The deferred tax working? And why. Why are we doing this and that? And, and even if it is at the point where it's going to. I think it's actually going to be a while, a little while, until it really understands all the integrations and be able to beautifully forecast. But even then. Even then, right, Giles, you're still going to have to understand what it's doing perfectly and if it's making a mistake, like a plug, be able to troubleshoot. So anyway, fascinating to see that. I hope that was helpful.

And I mean, overall, it's still like Elkhart, Great job. It's a great tool. We're not, we're not just out here being negative about this. Left, right and center. Obviously there's gaps, but it's still amazing what a tool like this can do. And I feel like we're saying that with, with all of them at the moment. Moment, which is interesting. So sorry, Paul. We didn't do anything too dodgy. None of us swore. I think we've held up the reputation of your podcast.

We even had fun without him, didn't we?

Oh, yeah. Not. We can't tell him we had too much fun. He won't come back, I hope.

No, do not tell him. We had more fun than if he was here.

I mean, we had him just slightly less fun than if he would have been here, but we still had lots of fun. That's what we'll say.

But you were rolling over a couple times with the, with things like the plug. But anyway, it'll be better when he's back. In the meantime, that was a great episode, a lot of fun and really hopefully interesting to see where this is headed.

All right, thanks everyone for watching, listening, and we'll see you next time. Bye bye.

Thanks, everyone. See you next time. Bye bye.