The Reality of AI Excel Tools for Finance Teams to Understand Formula Complexity with Ian and 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.

Today we're going to test Melder.

Done the first one perfectly right. I can see that, Charles.

It's choosing this substitute the dash with blank repeated 100 times. I don't know why it's doing it.

The M extract function extracts specific information from text or files using AI. So upload a PDF, say M extract everything about something. I could see where that could be useful.

And also 4 out of 5 bonuses. I know we don't normally check them, but it did it so quickly. And once we got over the process error, it did it in about four minutes, which I'd say is probably the second fastest.

Fascinated that a couple tools we've looked at have chosen that same technique over engineering to expand and then contract to come back.

One thing that jumped out to me is.

Welcome back to the MOD Squad. This week I have returned as one of the co hosts after Giles hijacked my stream yard. I've got back control and we're going to do another episode today. So I'll go ahead and introduce myself and then we'll introduce the co host. I'm Paul Barnhurst. Many people know me as the FBA guy. I'm lucky enough to have been hosting this podcast for a couple years and love the series we're doing on AI. We have an exciting tool we're going to test today, so we're going to continue with our testing of these different tools. And that's a little bit about me. Over to you, Giles.

Yeah, Hello, Giles. Mail. I am co founder of Full Stack Modeler and help people get better at Excel Financial Modeling. And yeah, really enjoying the series. It's been fascinating. I'm continuing to learn a lot and I feel like it's kind of. Yeah. Helping us all kind of stay up to date with what on earth is going on in this AI world. So happy to be here.

All righty in. Would you like to introduce yourself?

Thanks, Paul. I hope we didn't embarrass you too badly last week when you couldn't make it and you left the keys to the Porsche with me and Giles to take Dad's car out.

I thought you guys did a good job. I enjoyed the episode.

I thought we did, too. And I thought it was a very telling episode, maybe one of the best. And it would have been better if you were there, Paul, but it's still. It Was it was only because. And we'll see what happens this week. You know, we've, as I'm executive director of Financial Modeling Institute, my life's been involved with in modeling and it's been fascinating to see the trendlining the evolution of these tools. And you know, last week's episode in particular showed us all the highs and all the lows, all the greats, but all it. We got through every major pitfall and we'll see if we, you know, encounter similar issues in, in today's episode. So. Looking forward to it.

Yeah, no, I'm really excited. And so I believe we're on tool number seven now in our testing. So anyone wants to go back and see the others? We've tested Rosie, which has gone out of business. We've tested Excel, Agent Trace Light Tab, AI Truffle Pig, which is an AI spreadsheet. So that was an interesting one because it was not an Excel based. And then lcar. And today we're going to test Meldor. So why don't I pull up the screen. We can kind of look at the website for Meldart. But as we're doing that, so something that's interesting. I remember when we first started this and I looked at Meldr, it talked a lot more about how it had built its own custom formulas to extract data. Called it M extract and stuff. Now I look at the website, you don't see a single mention of that. And what I wonder is many of their formulas were like copilot and the copilot function came out and now they've kind of shifted. So it's interesting, these early companies, right, you see that product market fit of who are we Interesting. I would love if we had the machine, you know, kind of that way back and we could look at the website when it first came up and what we're seeing now because it's quite a shift. All right, so let's run through what they have. First thoughts we'll go to in first. Anything stand out on the website?

I feel like I keep saying the same thing.

I guess I know that's why it's.

Gonna say good luck on this because they, they all. There must be like, you know, a playbook for how to build websites for these. I like them. I like that they're clean, they're. They're succinct, they're clear. I'm not sure why. Like, how come when you look at most other websites, they're not clear and clean like this? This resonates well with me. But they're all very similar and if you scroll down, you'll see they all have. They all have a screenshot. Right. A couple screenshots. They tell you what they do. There's a bit of a sales.

They're.

They're vertical. They're a lot of. There's a lot of white space. Right. Giles? On these websites, typically in a good way. And that I find attractive. And I'm, you know, like all the movement and I find it a great presentable professional board. Do you guys.

Yeah, I suspect they're. Would it be wrong to say they're built with an AI templating website tool? That maybe that's why they all.

Look, I was just gonna say something similar. You covered what I was.

I like it. It's clean. Lots of Excel green. That's good. Good marketing tactics in there, I guess. No, totally outrageous statements. What did it say at the top? Your Excel AI assistant or AIXL analyst?

Yep.

Yeah.

I.

Again, you mentioned. What did you say plateau the other week, Ian and yeah, I feel like we kind of know what we're going to expect at the moment.

Couple of things I thought were interesting. Then we'll jump into testing. You know, built by former Google, Bain and PwC. Right. They're really leading into those names versus some of the others will say use by. Which means they're, you know, probably pretty early. It's backed by Y comnator. And then I thought this was interesting. Never write a formula again. Never make a chart again. Never clean data again. Right. Bold statement to say never.

Oh, I missed that. So they are bold statements. Yeah.

Yeah. But you know what?

I like those bold statements. I don't have a problem with that.

Because I don't have a problem either. It's still.

What it's saying is I like it, it's bold, it's confident and I'm looking over at it. But it is basically saying, I'm going to be your partner, I'll be your. To me. I read this as I'm going to be your sidekick. I. I'm going to do the dirty work for you. I'm going to do the grunt work for you so you can do, you know, higher quality things. It's not suggesting I'm going to replace you, which is some of them have almost implied that you won't be needed. I have no issue with this posturing.

Two other things and then we'll go. If you watch this one here, look at the formula on the screen.

Does that look brutal?

Four substitutes in a row.

Yeah. Giles and I Had a great bunch of laughs. And this is, this is the thing. So again it says. Here's the thing.

It's.

You scroll up again it says never.

Right.

And if you go up a little bit, never clean data. Can you roll up, never make a chart. Go with the one that says never build a formula again. Right. So I don't have a problem with that. Never built write a formula again. But I think some people will read into that. Oh, I guess I don't need to understand formulas ever again.

That's kind of a little bit of what I. When I see it I think of.

Right. So I don't mind if no one has to write them. But you still need to understand them because the one you looked at and we'll see what it does today in this test because if you can't understand what it's writing and we know that they're incline to write things that are complex and often verbose more so than person would do. If you can't understand it, you're in big trouble. So yeah, sure, never write it. Great. But you somehow still need to find a way to understand it.

The other thing I thought was interesting if you watch this last video here is you can see they have some of their own custom formulas.

It's.

It's called M for Melder. They have an extract one that extracts data from a PDF and I always.

Have an initial like defensive position when I see that they've got custom formulas because it's just more stuff to learn. That's just me. But as soon as I see custom formulas I'm like, oh God.

I guess for me it hasn't bothered me in the sense of my background coming up from FP&A. I worked a lot with TM1 and TM1 had a bunch of custom formulas to write stuff back to the database and read it. So I had to work with them all the time. That DBRW was the main one DB read write formula. I don't think a lot of it. But yeah, if they have a bunch, I think they have like four, three or four. We can. But you know, I agree with you. There's like 30, 40 custom. Like why do I got to learn all these? Now I got lambda. It just feels overwhelming.

And I'm not saying it's objectively wrong. It's just my. It's me being honest about whenever I.

Your initial response, like one more thing. I get it.

Yeah.

All right, let me close this and why don't we move into testing. Giles, if you want to Screen is up.

Hopefully you can see it. Okay, so I think we know what we're looking at here, but I will do a quick recap. This is what we would consider an entry level esports case for the World XL World Championship. You have seven levels, five bonus questions and they get harder in difficulty. Level one is very straightforward in one sense. We've got a text string. We've been given a number of characters we want to extract. You've got to fill out these green cells with a formula. We can been very specific. We want a formula to solve these questions.

That was one of the funniest excel moments of my career. And I. I imagine you've seen it, Paul, but wow, that was a great teenager cheat code. I even had a couple people comment to me on that from last week. It was hilarious. Let's see if Melder treats it similarly.

You're referring to the point to the answers moment, are you?

Yes, I. I love that when it.

Discovered the answer was there, it says that I'm just going to link directly to the an is brilliant. But let's see how this performs. Right.

It's how I solved my math problems in high school. I just went to the back of the book because we had to solve the odd ones and the answer was there and just wrote I on all the answers. I shouldn't admit that. Then I guessed on the tests and wondered why I got Cs and I think a D in my high school algebra class.

So I've hit. I hit go. It's starting to pop up with its thinking. It's just going to solve for level one. We'll see very quickly, I guess, whether it's so by lcard finished.

So, you know, when I talk about plateauing a little bit, I think that we would agree that capability wise, they've all performed well. Where we have been surprised. Correct me if you feel differently. Where we've been surprised a little on a couple of the tools more recently is less about their capability and more on their speed. Like we were just about to hit pause last week when the darn thing was finished and we're like, oh my gosh. Because instead of. So instead of taking five minutes as other tools did or seven, it took like 45 seconds. Right. So this one.

Yeah, it's thinking. It's delegating. Who's it delegating to? Oh, delegating to itself.

I love when AI tools delegate to other multiple personalities.

Oh, here you go. This is.

You have a manual or an auto approve. We have it on manual so we can. You're gonna have to approve it every time it adds them because you have manual down there. You see that down at the bottom in the chat box. Down below a little lower they have that manual versus auto. Do we know change it to auto and then it just automatically puts the answers in. In the interest of time.

Do we know what brain it's using, what engine it's using to. Do we know how. How the back end is working?

It doesn't tell us.

Yeah, I'm not sure. I don't want to click. Oh no, this don't go there. Not sure.

I tried clicking on its documentation and it said the page not found. So there's an improvement opportunity.

Anyway, it's done the first one perfectly right. I can see that, Giles. The first one it solved perfectly. No problem. And now you've run for the subsequent ones. Is that right?

Yeah. I mean again, by the looks of this, is that okay solution planning? I thought it had just finished all of the levels which would again would have been similar to LCAR speeds. Oh, something went wrong still. I don't know what that means. Okay, so try again.

Yeah, I'm not sure what that message error message is, but. So I don't know which model is it using? It doesn't tell us. Let's see. They did complete our survey.

I'll.

I'll look it up while it's running.

And see and you know, I have to tell you that there's a part of me that there's a part of me that appreciates that. I don't know, maybe it's just me being old school. I appreciate that. I'm. I have a piece of software and it's going to use the best engine capability of. I found it daunting. And there's a few tools that say now there's more and more, you know, engines you can pick from. And I just think that if knowing myself, I would probably run the same analysis using every one or most of them to see what. Which one's best and. And they keep.

I think with Elkar, I noticed Haiku was the one that was running, but there were other options. You just don't think I agree with you. There's something to be said for just tell me what's best.

Don't make right. Yeah.

So this is interesting. I don't know if you remember with Elkar we had that funny thing with Rept and 100 and this has done the same thing. And I still don't know what it's doing, but it's picked that same quite weird approach.

I mean it's chosen to repeat us. So again, first of all, again, can you click on that formula again? I mean this is just continues to reinforce our point. Good luck to anyone trying to. Good luck to anyone going to a meeting and having your boss say what exactly is this formula doing? If you didn't build it, didn't look at it it and have no idea what it's working. It's not. And none of these. I think we. None of these concepts in that formula are difficult. Right. The challenge is the embedding. Embedding them within each other. I don't know if you see it. It was. I forget what was in there. That long calculation, that long formula that it ran.

Right.

A large function is simple. A value function. I'm starting in the beginning. A value function is simple. The trim function is simple. The mid function is simple. Substitute is simple. The challenge is when they're all embedded and. And used as sort of a. Incorporated within one another to solve something. There's a lot of things happening. So in aggregate this becomes fairly complex. So why is it repeating a space a hundred times within a substitute function within the mid with it? Right.

Yeah. I don't know.

We would just have to piece it apart and we would all do it. But yeah, anyway, we might want.

Oh, hang on. Completed. Okay. So it's saying does it appear that.

It got it right? From what you can tell.

Giles, let me have a look at the answer sheet which is on a separate tab.

Workbook state is empty. Interesting. So while you're looking that up.

Yes.

Here's what they say about the LLM model. So I'll read this. I think it's interesting from this is their answer. We use Google Gemini models for the AI custom functions. They have four of their own functions which we can show a little later. These have the appropriate latency, blah, blah, blah to work. And then we use a variety of anthropic models for the actual agent, you're saying.

That's right. Is this on the. Is this on their website or is this in a response to you?

This is in their response. Remember we hadn't do the survey. This is what they said in their survey.

So I'm just saying this is not some information that in theory anyone would see on the website. They've told.

I didn't see it anywhere on the website. Maybe in their document. I'm sure they have some documentation somewhere that may give some of that information or. I would think they would. But yeah, it's not something.

Sure.

Do we want to pause for a sec? So it's. It's not. It's telling me it's got the answers but I can't see them. So we might just have to give it a bit of troubleshooting here, which.

All right, we're going to pause for a minute while we troubleshoot, see if we can figure out what's going on. And we'll be right back. We are back. After technical difficulties, we had to rerun it.

So why don't you take us through.

What happened here, Giles?

Yeah, it. It seemed to just error out a couple of times, but I shut down the workbook, reopened it, reprompted the same thing and I think we thought four minutes and it's actually this time flown through. So just like in the last episode I was on, it's choosing this substitute the dash with blank repeated a hundred times. Genuinely. I mean, there might be a really logical answer to that. I don't know why it's doing it. Level three, column indirect is correct. I don't know why it's adding a one to the row number. That is completely unnecessary. But it's done it. Len minus Len substitute when we're counting the characters is fine. That's the right answer. Then in level five, what's. So this was a very similar one. And again, I don't know what. It's just so interesting that more than one tool has turned to this rect a hundred times. Solution. Level six, indirect on the map. Correct. Level seven, indirect with movements. I'm assuming there is a. There an offset. Oh, interesting. It's done it without an offset. Oh, it's adding a row.

I think it's using address row and indirect versus an offset.

Yeah. So it's. It's. I think. And also four out of five bonuses. I know we don't normally check them, but it did it so quickly. So that is almost a perfect score, which is as good as any. The best that we've seen. And once we got over the process error, it did it in about four minutes, which I'd say is probably the second fastest.

Yeah. Toward the faster end. So it did well. Once we got past the issue the screen, I don't see any kind of to do list. It seems very. Which I like. The pain seems very basic, but you can follow everything. Like it. It feels a little more text heavy than some of the others, but yeah.

It definitely showed a lot of its thought process. You're right. I don't. I don't see checklist Although we do have a closing checklist of what it's done. Uh, I. I think that's a really good performance, to be honest.

Yeah. Why don't we. Should we move on to the next case and give it a try, and if it bombs out, we'll just move to the next thing?

Yeah, fine. Okay. We'll do it really quickly, just in the interest of time. So if only just to see my face on a spectrum.

That's really why we want the humble mvp, is to see Giles face only.

Just for that pleasure, let's go over to my prompt. So we're going to get it to autosave. Needed. We strongly recommend you autosave.

Oh.

Oh, your free trial has ended. Oh, that's Elkar. Wrong. Got so many tools on my add Ins section.

I swear, you get a couple days.

Look at you, the richest.

Use it for one exercise. Get out of here.

One of them I can't delete at the moment, and it's got access to all my data.

Oh, minor details, minor details.

So that has instantly said failed to build Agent Focus. This is like another naughty teenage analyst. It's like, well, sort it out. I want you to focus. I've given you the time.

Now, in fairness, I know they do say there's. They do have some pricing. They still say kind of. They're in an open beta. They're still working some things out. But this is concerning how much we're getting errors here. Right.

Not a. I'll try reoccupating time and let me see if I can get it to work just to give it a chance. And again, this isn't like an objective. Oh, my God, it's broken. It's. We just do see this a lot where sometimes these tools bug out and just won't work. So try one more time. Here's the prompt. It's. Oh, this time it works. I don't know. No, it doesn't matter. Okay. Failed to build Agent Focus. I think we just have to say we can't get it to run the same thing.

Refresh once and see if it makes any difference up at the top.

That just.

I just dropped it. Yeah. Okay. Once it fails, we'll just be done reading workbook.

Read the case levels. I think it's just seeing my face and going, no, I don't want to do it. Don't want to do it.

All right, well, why don't we move on? Probably the trial balance. Since this one. We'll just move past this one.

Yeah. Okay, so let's go. Let Me bring this up really quickly.

And you know, in fairness to Melder is, as they said, look, we're still in beta, they're working through some bugs. So these things are going to happen. Doesn't mean they may not be able to perform. It could do great, but just the reality of testing.

The reality of testing. Okay, so let me hit the prompt. Remember, this was a raw trial balance data in a tab of Excel. I'm asking it to look at the raw data and essentially produce some sort of an output. A tab with some KPIs with some visuals. So it is having a go at.

That now and it appears to be solving this one. So that's a good sign. Do we want to pause it here for a minute or while it's running? I'm going to share my screen for a minute while it's running Giles and just show something I think is unique to the tool.

So I think if I pull this over, we are getting a lot of information about its thought process and what it's trying to do, which I think is good.

I mean, modifying and. All right, so we'll let it work for a minute and we'll come back. I'm going to share something else here. So I think it is interesting. If you click on Help, there are two things. It lists the custom formulas. That has an example. So MGEN is basically Copilot. So it's just a generate. They. They'd come out with a. Before Copilot, you could see, put your prompt where you want it. It's basically Copilot. Then they have this one that I think is interesting. The M extract function extracts specific information from text or files using AI. So upload a PDF, say M extract everything about something. I could see where that could be useful, right? Uploading a file. So kind of interesting. You know, extract the vendor name, the candidate's most recent employer, whatever. Then they have one that's specifically designed, which you could do with, you know, Copilot shows you to do it. But it has one called Categorize. So you use it to function lets you categorize one item into a preset list of options. So the example they give is New York, rural, suburban, urban. So identify which. Which risk categories the contract should fall under. And then they have one that's supposed to work for web scraping. Says not every website can be scraped. So keep that in mind. But so interesting. And then, you know, it has the uploading of files. So upload PDFs and viewing files and citations. So I'll bring it back to you and how it's performing, but I just thought that was interesting because it's the only one. They said, hey, we're going to try to give you some formulas to do certain tasks.

Yeah, okay.

All right, Let me stop sharing and we'll bring Giles back up while it's running. Go ahead.

So, I mean, it's still working through it again. What. What I do like is we're getting a lot of information. What? I don't. Oh, that's interesting. So it's not created a new tab. It's overwriting. The monthly summary tab was there, so.

Don't want that.

Yeah, you definitely don't. Oh, you don't.

Wow.

Yeah. Okay. I think we pause. Let's see how it gets on. But I'm a little bit nervous that it's overwritten a summary that was already there, which you wouldn't really want.

Agreed. All right, so we'll pause it and we'll be back here in a moment to see what we get. We're back. Looks like it's still working. How's it coming so far? It's been running a while.

It's been running a while. What's it been, like, 10 minutes? Maybe a little bit more. I mean, I can click in kind of. So I think it's getting there. It was very odd that it decided to overwrite on an existing tab, which is not ideal, but I suspect if we left it with a lot more time, it would get there. I mean, positives, it is looking at the trial balance data. Some of them haven't even done that. The formatting isn't there yet. I've got some div. Zeros. I mean, it's coming out with strengths, concerns, recommendations, financial performance. That's pretty good. Yeah. I think it seems to be doing okay. It's just taking a very so well in the first world problems of 2025. It's taking a long time, I. E. 10 minutes.

Yeah, it's been running for a little while, so why don't you just let it keep running? And what we're going to do, we thought this would be really valuable. As you remember, it solved that Excel esports case with that really long formula with the rep 100 and the mid and the substitute, and we're all kind of like, what is it doing? Well, we've been here testing in, took a few minutes and broke it apart. And I think this would be a really good learning opportunity here to show how it's important that you can take formulas apart and understand. Is this over? Engineered? What are they doing? Does it make sense? So why don't we bring on in screen and we're just going to talk through this a little bit and show you what we did, what the tool did, kind of how to think about it. So in take it away.

Sure. And you know, thanks, guys, and I think this could be a very helpful learning, you know, opportunity here. In my past life when I was running a training business and I taught intermediate and advanced Excel programs all over the world, you know, I would always tell people there's nothing complicated about any particular Excel function if you dissect it. The large function is simple. The value function is simple. Trim is simple. These functions are very simple all by themselves. Excel gets complicated when you start nesting and embedding functions within functions. And you know, I've always warned people that just because a person or an AI tool is doing this does not necessarily mean it's the best and most efficient way. And you know, what I keep preaching and talking about is if you're allowing AI to do this, you're going to need to understand it because there might be a simpler solution. Now, I have not picked apart all of this to know why it's doing this. It seems overly built. But in a nutshell, what it did is it. We asked it the question said, hey, there's a column of, I just take out one of them here, there's a column and they all look like this that have 10 sequences of digits, 1, 2, 3, 4, 5, 6, 7, 9, 10. And the question simply says, break it apart and add up the largest and the second largest value. So in this case, the 89 is clearly visually the largest. 79 is the second largest. Add the two up, you get 168. So it did that by this, you know, extremely long formula. Is it the simplest way? Well, let's see. What I did is I started by taking, I pulled apart and this is a good practice. I pulled out the substitute portion. I wanted to understand why it was doing this, why it was looking at the cell reference, taking the, the, the dash and repeating it with a hundred spaces. That's what it's doing. So I pulled that into its own cell and you can see it's literally looking at cell F5. It says, hey, anytime you see a dash in this one cell, replace, replace that one dash with a space and not one space, a hundred instances of the space. So now instead of having a single dash, we have the number 35 followed by a hundred spaces, then 79, 100 spaces, et cetera. Now we'll have to go through and understand again and then it pushing them back together with this mid function and it's squeezing them back. But I thought let's see if we could do this much simpler. Well, the easiest way to kind of get to this would say, would in my mind would be to say, let's just look at the text split feature. The text split function just says what cell do you want to split? And then what's the delimiter? What should you look at splitting it? And I'm just going to say here, anytime you see a space and then a dash, let's split it there. Well, that's easy. So now what I have is I literally just took this one cell and I split it into ten columns. That's easy. And I thought, well, now that I've got them into separate cells, why don't I just see if I can figure out the largest value. Now this is an array. This is a, this is an array. This is a spilled array. So what it does is it gives you, when you look at a range, it'll give you this first cell reference and then a number sign or you know, a hash at the end. That's fine. But you'll notice it's not quite working. It's not giving me the max value should be 89. It's not working because it doesn't appreciate that these are numbers. Okay, so what I realize I need to do, ah, is I need to go to my text split and the value function just says force it to recognize that these are actually numbers. Okay, good. So I've said, yeah, this whole range of cells, they're numbers and the value function forces it to recognize that.

Good.

Well now that I've done that, the max function looks at this whole column and says the maximum value is 89. The second largest I could figure out by using a large function which says look at the range and then put a two at the end. So now it easily knows that the largest is 89. The second largest is 79. I mean I could just add them up with a sum function. So literally by adding a few extra rows or columns, I could have replaced all of this with a text split and with a value, a max.

Oops.

And then a large, a large and then a sum. Now if I wanted to get fancy and do it all together, let's do it one more time for two seconds and then see, I could once again do a text split on this range and again say let's do it. And I want you to see Once you understand what you're doing, it's not a big deal. It did it. But one more time, let's make it into. Let's make sure it knows that there are values and then what I'm going to do is say let's take the maximum value within that array and then what if I add that together? What if I take the whole thing again and repeat it and now say take the largest one within that array and let me find my cursor here and take the large function and put a two and the second largest and there I have it. So in one, very, very simple. This is about a quarter of the length but now I understand it and this is kind of the process we want people to make sure that they can understand as opposed to simply relying on something like this that might or may not be correct and may not be understandable. So anyway, I thought that was interesting and I was doing that when you guys were chatting because I was convinced. Convinced that this is not the most.

Efficient way to solve it.

So Giles, back to you.

Giles, you looked up something on the rept, the repeat function.

Yeah. So the. I put it in. I just put the formula in Claude and it said because it's using mid, it's it's future proofing or mitigating against potentially having the mid cross over into the next digit. So it's putting more space than you need because the mid might then cross over into the next number. So it feels over engineered. Apparently it is a method of doing it. If you share my screen again, I just thought I'd show you that a.

Couple tools we've looked at have chosen that same technique over engineering to expand and then contract to come back when I believe that it clearly wasn't necessary.

Here you go. This is sort of the sort of stuff that I pick up from the esports. So you can use large and you can just use the curly brackets to pick the two largest digits. So this would probably be what I would do to just. That's probably as short as I can make it to get it down to that.

Look at that. That's even shorter than what I just did. I did it my way so people understand it. This is a little bit more obscure but, but, but even shorter. Right. And it works great. But the point is look at what you did. Can you click on your solution? Like just see how many characters that is in total? Like 15 characters in total. Now click on the green one.

Right.

So again, great. Oh no, sorry you. That's what AI did.

Right.

This Is Mel. Yeah. So I think this was the second run. So in the. Which again is interesting, in the second run it did it differently.

It did it differently. So the first one was what I showed, which was a monster formula there.

Whereas that's interesting. So it's done a different.

I just continue to believe different ways for different rows.

Yeah, I continue to believe that people are going to have to elevate their skills to understand. It's great. You're going to learn a lot of things. You're going to understand features you maybe haven't seen before, but you have to understand. So.

So back to just round off my test quickly. Bizarrely, it got a long way through, built loads of stuff and then it deleted all of its stuff and it said, right, I'll redo that. And it started and then it's broken. So I think at that point we should just admit defeat on this task.

We'll move on from here. What we're finding is that, you know, like I said, it's in beta. It's clear they have some work to do. But we could see, similar to others, they did quite a bit of work there and it appeared to be good work. So it's on the right track. But we're not seeing anything here outside of hey, it has some custom formulas, PDFs. There's nothing really all that different. It just has development to do. Kind of think that's a fair. Did really well on the first one. All right, why don't I share one of my use cases and then we'll move on to INS here. So I'm going to do the pvm. We'll do that kind of quickly here and see if we run into any issues. Let me pull up my prompt. We'll drop it in here.

This one should only take it a few seconds. It's usually pretty quick at this. I think we've got to the point failed to build Agent Focus. All right, let me try creating a new chat. We'll hit reload and we'll try one more time. If this doesn't work, we'll move on to modeling and if we have problems, we may have a end this episode a little early.

Or you could just do a little song and a dance. Paul. I think people would like that as well.

Something went wrong. Interesting.

I'm just going to type.

Can you calculate and see what it does? The price variance mix on this worksheet for each product, put the answers in the yellow highlighted cells. If I can type, I can't type. In case anyone was Wondering in column N and just see if it even failed to build Agent Focus. All right.

That's what I had.

I'm going to try closing that, reopening it once, and if it doesn't work, we'll move on.

Yeah. And like you said, it's in beta, so I'm quite. I think that's right, isn't it? This tool is in beta, so I have an element of kind of support and I feel forgiving of it.

Yeah, it's an open beta.

Yeah, I knew.

I know. They do list some pricing so that. I don't know. You always.

Paul, you always seem to have an issue when a company is in beta but does still sell their product, don't you?

Yeah, I don't. They say they're in beta, but their pricing, their pro is $50, which is one of the highest of any of them.

Okay.

Well, I could tell you based on what I'm seeing in beta right now, I would not spend that.

Yeah.

Right. So definitely, they. They have some issues to work out if it there. And maybe it's just today, maybe 99 of the time it's up, I don't know. But all we can do is what they're go with what they show us. Right. So we'll try one more time. Let me find the right prompt. Give me one moment here, people. I got too many browsers open. All right. Got my prompts. Try this one more time. I agree. If you're in your beta and you're charging people, we should be able to test it. And I expect to see a product that's providing that value, at least the value of what you're charging. All right. Failed to build Agent Focus, so why.

Don'T we move on?

Move on. I'll try to see. We'll try the deferred revenue schedule, see if we have any luck there. And if not, we'll try one of IN's exercises. If that doesn't work, we'll just do a song and dance for the rest of the episode. Brought to you by Giles. Giles will rap for us.

He does have some good rapping skills. You're gonna. We're gonna pulling to the bottom of the bag here, aren't we? But Giles. A rap from Giles, I think would be the highlight of any episode.

Yeah. Yeah, that would be interesting.

Maybe we can beatbox behind him to see to. To set a beat. Can you that, Paul? We can get a.

We could ask AI to do that, I think.

Okay. We'll get a beat for Giles. Do you have your Glasses handy. Giles, could you. Paul, look at that. The. The glasses are right there if he ever needs them. A costume change.

Needs that fur coat. But I know he's upgraded now He's. He's bigger than just a rap.

Is it going? Is it working?

All right, well, let me. Let me kick it off and see what happens here.

I can remember the beginning of my lambda wrap, but then.

All right, go ahead and get us started while this is running with a little lambda wrap.

I just have a. I have a beat in my head. I took my eye off Excel for a week or two, came back. Now it does stuff I never knew. Cells are spilling, arrays are going wild. Formulas stretch out like they've somehow been restyled. There's a sort. A filter. There's this thing called let. Some function called scan. I haven't learned that yet. I'll carry on. Done. There you go. There's a sequence running and a map on.

Cool.

And I'm just sitting here scrolling, trying to recall. I swear I was gone for like one long nap. And now Excel feels like a massive coding trap. I've got questions. They stack like rows, but one keeps punching me right in the nose. I used to sum cells. Now it feels like black magic. And the help docs read like a logical schematic. Tried to run a calc and it filled me with anger. So tell me one thing. What the is a lambda boom?

Boom.

He's gonna put us out of business.

Paul.

Paul. He. We are. We are very stinked. We're gonna be.

I agree. We are in trouble. It's working now. So it looks like the deferred revenue schedule. So I don't know if there's something in certain workbooks. This ca. No idea what's causing it, but we're getting the occasional error. Why don't we pause here? It's giving me options. Okay.

I understand the issue. Let me try a different approach. So it came into an issue. Since you specifically requested this be in a new worksheet, please create a new blank worksheet in your workbook. It won't create it for me. Once you've created it, please let me know the name of the new sheet and I'll build the complete deferred revenue schedule there with all the monthly columns and formulas. Alternative, you'd like me to place the schedule in a different area of the raw billing data sheet. So it can't build it. It can't create its own sheet.

So that's why. Yeah. It overwrote my tab data. Yeah.

Starting Seven columns to the right. I can see the Rob building.

It can't.

What?

What can't it do?

Create a sheet. It's saying I have to create the sheet and then tell it to put it on the sheet. So here's what it recommends. It gave me two recommendations. So I recommend you manually create a new worksheet. So right click. Okay, I'm going to right click here because I can't hit the plus button. It says to right click. We're following its directions. Insert or press shift F11. Name it something like deferred revenue schedule. All right. It didn't tell me how to name it, though. It didn't tell me to do rename. But revenue schedule. Then let me know the exact name of the new sheet and I'll build the complete schedule there. Please build the complete schedule on the worksheet. Deferred revenue schedule. Interesting. I was that delivered.

I don't.

I don't understand the reasoning behind that. I mean, I think that would have to be a deliberate choice.

Yeah, it must be. Well, that's the first time we've seen that, I guess.

Thoughts on that, Giles? I mean, do you like it?

It's just a lim. The fact that it can't build its own sheet.

Yeah, that. It tells you you have to name the sheet. What's your thoughts on that?

If I'm using AI, I want it to do it for me. That's the whole point. Right. So it feels like a bit of an odd limiting factor.

It does. I'm trying to think if there would be any security, like reasons of forcing you to make sure. So it's the name you want. I can't think of any good reason. It's interesting. Here we go again. Hard coding. Those dates. They love that.

Do we want to pause? I have a feeling this one might take a little while.

Yeah, let's go ahead and pause and give it a minute. All right, we're back. And we got it. To build the deferred revenue schedule, I. I had to create the sheet which we talked about. And you can see, you know, similar to all the others. It gave us instructions, key features built. It gave a few examples, saying, hey, they tie out. And similar to all the others, you know, put in these first few rows. What it did do. Is it linked to the front sheet. Fine. Brought those in. I would like. Some of them have done a lookup, which I prefer because if you change orders or anything, it just kind of keeps it all aligned. But the one thing that jumped out to me is here so two. Actually two things jumped out. First, again, hard coding dates. That seems to be something a lot of these tools will do if you're not very specific. So it gets back to being very clear in your prompting. You know, at least it did do a real date here. If you remember some of the earlier ones. We got things that weren't dates here. But the other thing it did is it linked back to the front worksheet, whatever row it was in. Right. So if I change the sort order here, this should not. Oh, yeah, it's gonna. If I change the sword order because it linked to the sort order. So that will work. But I don't like it all being linked to the front sheet. I'd much rather have lookup. So it works. But just a kind of an odd way to do it and say, hey, I'm going to take the months off the front sheet versus doing a formula validation. Fine, that works. So there we go. Similar to all the others. Took a little different approach. But what we're going to do now, we had an opportunity to talk for a minute given the tool still in open beta and we've seen some challenges, the errors. We're going to forego going through all the spreadsheet exercises. But what we do want to say is I think, you know, Melder has shown when it's working, it's similar to many other tools, has a few, few differences and a few limitations, like creating your own worksheet as a quirk that I'd like to see them change. I do like that you can do some web page. I like some of the custom formulas, I like that you can load your own files. Some of the other tools don't allow that. So it does have some unique features, but has some room for opportunity from my perspective. So any final thoughts here, Giles? And why don't we start with you, Giles?

Yeah, Again, that word plateaued is really in my mind at the moment. I think it's another reasonable tool. It's clearly got some issues. I think if it had overcome just the processing issue of not failing to do the prompt, it probably could have performed pretty well across the board. Have we learned anything new from a bird's eye view? No, I feel like they're all doing lots of interesting good and bad things. So, yeah, that's where I'm at.

The good that I like here is you didn't have to select the model. Not that, I mean, not good, but different. I think there's some benefit to that. It's kind of nice sometimes to not have to be thinking, oh, I should use this one or that one, or am I using the right model?

So, in final thoughts here, I believe this is only the second tool that we've tested that made the important upfront decision not to embed within Excel. Right. But to be built as a separate, independent.

No, no, this one is embedded in Excel.

This one is.

Yeah, I was, I'm sorry. I'm sorry. The next one we're. Next time we're going to be doing. That's right, sorry. This one. Yeah, this one was within Excel guide. So it was more similar. That's right. I'm thinking ahead. Uh, yeah, I mean, listen, we know that they all use similar, you know, brain power behind the scenes. Why did this one not quite perform where others have performed better? Not. No, not sure. But yeah, I mean, again, we. We've set a high bar. Right. We've set a high bar now where we have come to have expectations. It's still pretty wild that it can solve problems and, and do things when you ask it to, but I agree with you. I don't know that I would be using this for any, you know, critical, pivotal tasks at this point, even though it's fun to play with and fun to see, you know, how it performs and, and, and, and, and think through it. But anyway, interesting to see it and I guess they've got some work to do, like everyone and I look forward to seeing how they evolve and continue to improve over time. We're not here to, to be critical, but rather just to kind of be observant and to kind of to demonstrate how these tools are performing and, and hope that we can watch a continuous improvement and evolution as they grow.

Agree. So I think the final thought is we're excited to see how Melder develops. It has some unique features. The formulas, the way it thinks about files. It's still an open beta. It does charge, which, you know, that's why we're testing it. It's open for testing and we think there's definitely opportunity to grow here. Nothing unique we're seeing with this tool. We're excited for our next tool. We'll be testing for the next episode. We're going to be doing subset, and that one will be our second, as in revealed earlier, our second spreadsheet tool that is not in Excel at all. All the others have at least had an Excel application. All right, so we're going to go ahead and wrap this episode up and please reach out to us with any questions or thoughts you have. Thanks for joining the Mod Squad. And, Giles, thank you for wrapping this episode.

Pleasure. Didn't know I could do it. There you go. It was news to me.

Brilliant.

It was brilliant. I agree.

The My Squad.

We are the My Squad.