Testing The New Microsoft Excel Agent for Finance Pros with Ian and Giles

In this episode of The ModSquad on Financial Modeler's Corner, Paul Barnhurst, Ian Schnoor, and Giles Male explore Microsoft’s newly released Excel Agent, a beta tool designed to bring AI into Excel Online. The team compares its performance to Rosie AI, running both through a range of tasks including formula building, audit reviews, and a full five-year model forecast. Along the way, they test the tools against real modeling challenges like the Excel Esports case and a creative custom case, "The Humble MVP." This isn’t just about flashy tech. It’s a deeper conversation on where AI can help, where it falls short, and why core financial modeling skills still matter.

Expect to Learn

• What happened when Excel Agent tried to build a full five-year forecast model.

• Why solid modeling skills still beat "one-click" AI hype.

• Where AI tools succeed, and fail, in reviewing formulas and detecting issues.

• Why you should never fully trust AI without understanding what it’s doing.

Here are a few quotes from the episode:

• “Building a model is like flying a plane; you don’t want a rookie in charge when something breaks.” - Ian Schnoor

• “You can’t hit a prompt, go get a coffee, and expect a working model.” - Giles Male

Follow Ian Schnoor:
LinkedIn -  https://www.linkedin.com/in/ianschnoor/

Follow Giles Male:
LinkedIn -  https://www.linkedin.com/in/giles-male-30643b15/

In today’s episode:

[02:21] - What is an Excel Agent and why does it matter?

[05:37] - Sharpen Your Modeling Skills

[06:51] - Modeling Requires Human Judgment

[09:27] - Esports Case: Rosie AI vs Excel Agent Test

[19:39] - Excel Agent Hangs in Beta Mode

[25:26] - Testing PVM in the Smoothie Business Case

[30:01] - Ian Takes Over: Testing Excel Agent on AFM Model

[44:11] - Building a five-year forecast model from scratch

[55:19] - Adding a revolver to the model

[01:00:12] - Why the balance sheet doesn’t balance

[01:03:46] - Final thoughts and takeaways

Full Show Transcript

[00:01:05.34] Host: Paul Barnhurst: Welcome to another episode of the ModSquad Pod. I'm your host, Paul Barnhurst, aka the FP&A Guy, and I have here with me my other two hosts really excited to do another episode. So why don't we start with Ian? Why don't you give a real quick introduction to our audience?

[00:01:22.30] Co-host 1: Ian Schnoor : Thank you Paul, it's great to be back here with you and Giles, and nice to see the rest of you watching. I’m Ian Schnoor, executive director of the Financial Modeling Institute. I'm running our tests on three statement modeling, corporate finance, modeling. And we'll get into some other tests like that applicable to those working in sort of accounting and finance, corporate finance fields. My whole career has been in financial modeling, and I'm excited to be working with the two of you on this great podcast. Over to you, Giles.

[00:01:49.70] Host: Paul Barnhurst: Why don't you do a quick introduction of the humble MVP?

[00:01:52.70] Co-host 2: Giles Male: Sure. Well, it's a pleasure to be back again. And I'm particularly excited about this episode. Actually, with all the stuff that's gone on in the last week online. So the co-founder of Full Stack Modeler, Microsoft MVP, Lifelong Modeler, and esports, I should mention that more Excel esports player this year. Very average player.

[00:02:12.20] Co-host 1: Ian Schnoor : No, you are above average. You're scoring well.

[00:02:14.64] Host: Paul Barnhurst: He is above average.

[00:02:15.64] Co-host 1: Ian Schnoor : He's humble.

[00:02:16.32] Host: Paul Barnhurst: That's why he's the humble MVP. He won't let his real.

[00:02:18.88] Co-host 1: Ian Schnoor : Score an outstanding competitor. Yes.

[00:02:21.40] Host: Paul Barnhurst: So this week we're going to dig into Excel agents. About a week ago, maybe a week and a half ago it wasn't very long. They released the beta version of the Excel agent which you have to have a copilot to access. It's an online tool only, but it does many of the things similar to these other tools. And so it's got a ton of publicity online. I know, Giles, you've had many conversations, seen some things, kind of maybe share a little bit of your thoughts so far with what's going on as far as Excel agent.

[00:02:50.84] Co-host 2: Giles Male: I asked a question on the Unpivot podcast months ago saying, I wonder what's going to happen if Microsoft, like, catches up. So the Excel agent for me, I'm really excited about. I think it's fantastic if Microsoft can lead the way but have pressure from good competitors, I'm happy. What I'm not happy about, which I may have expressed once or twice, is where people get a little bit carried away with the language and say, oh, you just click a button, get a coffee, and your model's built. And I think we need to really unpick whether that's true or not if it is awesome. But let's see.

[00:03:28.54] Host: Paul Barnhurst: So what are your initial thoughts? I know you've seen some things online. We've talked a little bit about this. Where are you at now that you've had a little time to think about this? I know we've touched on it last week.

[00:03:37.30] Co-host 1: Ian Schnoor : I don't tend to spend as much time, maybe as some other people online getting involved in some of the so.

[00:03:43.94] Host: Paul Barnhurst: It has more of a life that you and I do.

[00:03:46.30] Co-host 2: Giles Male: You're a mature adult.

[00:03:47.82] Co-host 1: Ian Schnoor : I'm not sure that's true. I'm not sure that everyone would agree with that statement, but maybe I'm just lucky that I don't have quite the time, but that's fine. I just spent some time online. Listen, I've seen what's going on. I don't love you knowing me. I think we're all in a similar boat. We. We're all excited about all of these tools. The Excel agent, while it was frustrating to get launched and installed for me and I couldn't get it working easily, and of course, right now it's only in the online cloud version of Excel and you seem to need a copilot license, so it's not the most straightforward path to get it. But I finally managed to get it working. It's impressive. I mean, listen, I will always be impressed when I can, you know, write an English sentence or presumably someone can write a sentence in any language and then you just watch it building things in a spreadsheet. I mean, that's still pretty mind blowing. It performed very, very well. But as you said, Giles, we're not talking about where you enter a sentence. It does a whole day of work for you and you go home. I mean, we'll see that in the tests, especially when it comes to model building. So but really impressive start. And I'm excited about what we're seeing here.

[00:04:53.11] Host: Paul Barnhurst: Yeah. And one thing I said over the weekend is we saw, you know, a lot of people getting really excited about it and maybe over-promising on what it could do. This is my conclusion. And it hasn't changed. I think it really was at the start. I'm more convinced of this now if you're not a good modeler, don't start with AI prompts. Go take a course, build some models, build a strong foundation, then use AI to help you if you're using it for assumptions. I like to help validate assumptions. Give you ideas of inflation, maybe some ideas for scenarios, how to write a formula. All those things are great, but don't just think I'm going to become a modeler today and write your prompt and turn it into somebody. That's just a disaster waiting to happen.

[00:05:37.95] Co-host 1: Ian Schnoor : Well, I'm going to show in my testing, I am going to show and I agree with you 100%. People need better modeling skills than ever, I believe. Take a course with one of you. Get your AFM certification right. That proves that you have incredible modeling skills because it is only at that point I was able to find errors in some of these AI generated models that others might not find unless they were also really strong modelers. And so. So it is important you do not want to be caught off guard with a mistake that you couldn't catch.

[00:06:04.53] Host: Paul Barnhurst: And that was the message from so many from my posts is, you know, Lance Rubin and many others are saying, I fully agree with you that it's going to make the best modelers are going to get benefit out of it. If you don't know what you're doing, then learn what you're doing. Giles, you're going to say something. I saw that little.

[00:06:22.13] Co-host 2: Giles Male: I was just going to try and pre counter the accusation that, like, for me personally, is the co-founder of a trading business for Excel and modeling. It would be like, oh, of course you would say that. I don't care where you get training. If what you just said Ian is spot on. Like if you want to go and train with anyone that's reputable, please do that rather than do what you just suggested, which would be to just go, oh AI, I'm a modeler, hit a prompt, go and get a coffee. Like I just get the training, you know.

[00:06:51.67] Co-host 1: Ian Schnoor : Building a model, as we all know it is a discipline and it's a skill. It requires judgment. It requires thinking and asking good questions. It's no different than any other skill like playing a sport or flying an airplane. I mean, in theory, in theory, I could be, you know, chosen to be a pilot of an airplane and 99% of the chance that if even though I know nothing about flying a plane, I could become a pilot tomorrow, and, you know, the plane could take off and land, and I'll get lucky because there are AI tools built in to, you know, airplanes and autopilots, that it'll work, but you don't want me there in the odd chance that that something goes wrong, right? That I need to troubleshoot. Um, you want to make sure that whoever is driving this engine, driving this machine knows how to troubleshoot, knows how to look you in the eye and assure you that it's all working, that they're all set. Um, because things do go wrong a lot, and you need to make sure that you can step in when needed.

[00:07:43.79] Host: Paul Barnhurst: It's a really good analogy. You think, right, commercial pilots fly those big planes. I think they have to have what, like 1000, 2000, so many thousands of hours before they can even get that job. And it's not because the tool doesn't do 80% of the flying. It does a lot of it's autopilot and the machine. But when those landing gear doesn't work, when turbulence hits, when the wind goes on fire. I want that experienced pilot.

[00:08:11.61] Co-host 1: Ian Schnoor : And by the way, I mean, you don't even need AI for airplanes. I mean, airplanes have had autopilot mode for a long time. Sometimes when I run webinars all over the world, at times I'll start and I'll, you know, ask, you know, the group or the class? Hey, who here would get on an airplane with no pilot? Imagine one day walking onto an airplane and they said proudly, look at this. You're going to fly and you're going to go up in the sky. There's no pilot. Um, now there are taxis doing that, of course, and there are cars. And maybe it's just a mindset and we will get there. But right now, I can promise you, most people I speak to would not be comfortable being on an airplane flying it if there was no pilot at the front. Even if they thought about it, they'd realize that when there is a pilot, that pilot may not be doing that much on their own anyway. But they're there. Right. And I don't know if you have a thought about that analogy, Giles, but that sort of feels relevant to me in what we're talking about.

[00:08:58.71] Co-host 2: Giles Male: I do not want you flying any plane I'm on any time soon. So.

[00:09:03.47] Host: Paul Barnhurst: So would you rather have nobody flying it? Is that what you're saying, Giles?

[00:09:06.87] Co-host 2: Giles Male: Oh, that. Well. That's different. I think I'd rather have Ian there than nobody. Just. Just as a backup. Uh, yeah. I'd take you over. Nobody. There you go. There's a compliment.

[00:09:14.43] Co-host 1: Ian Schnoor : And you know what? You probably would. Because you probably take a human because, you know, presumably you could speak to air traffic control, and they could try and give you some instructions. Right. But what you're saying is you want someone there who can help troubleshoot. Important.

[00:09:26.83] Host: Paul Barnhurst: All right. So I think we've covered some great points there. I want to share one other thing. I did some testing over the weekend. I showed this to you guys earlier and we thought we'd bring it on the show for a minute. So I'm going to share it. So let me just level set what I did and we'll bring it back again after we've done some more. But I took the easy Excel esports case that we've tested the last few weeks that the Excel athlete case and I ran it through Rosie AI and I ran it through the Excel agent, and I ran it through each of them twice. And so what we have here is say, what model do they use? This one's using a copilot. This one's using ChatGPT. Very similar. What's the level max score? How they scored e.g. haven't got the exact same score. They each got 1210 out of 1250. They missed the bonus question five. There's a very easy reason why they did. They both took roughly the same time to run it within a second of each other. They all use formulas each time. All of them were formula based, and in some cases they use the exact same formula between the tools. So let me sort this by level, because I think it will be kind of interesting. So right I could see oh hey, they use the exact same method for level one.

[00:10:38.37] Co-host 1: Ian Schnoor : So you are, this is Excel in the cloud. You're showing us I. Is that correct?

[00:10:43.79] Host: Paul Barnhurst: I'm using Google Sheets right now, so shame on me.

[00:10:46.11] Co-host 1: Ian Schnoor : Oh my goodness. Giles, close your ears.

[00:10:48.83] Co-host 2: Giles Male: It's fine. I'm relaxed about things when people do things that are slightly different to me, it's fine.

[00:10:54.59] Host: Paul Barnhurst: All right. So question one. They both answered it the exact same way. Question two similar but different approaches. Right. One put a value in the front one used the double negative in the front to convert it from a text string.

[00:11:11.23] Co-host 2: Giles Male: And I think.

[00:11:12.43] Host: Paul Barnhurst: One did a sum here. Like you mentioned before, one did a plus and did an array constant.

[00:11:16.75] Co-host 1: Ian Schnoor : So can you just remind us, can you go left for one second just to show everyone one more time? You're talking about these each of these rows just so we can see it.

[00:11:24.03] Host: Paul Barnhurst: Each row here is the first row is rosy. Second is Excel agent level one. In fact let me hide this.

[00:11:31.91] Co-host 1: Ian Schnoor : So. That's right. So just a reminder you're showing how you ran it under rosy AI, which is a tool we looked at last week versus the new Excel agent. And you're saying question one, they did a very similar, um, the same the same thing, didn't they?

[00:11:45.97] Host: Paul Barnhurst: You're saying the exact same thing. There's really kind of one simple answer for that. So they did the same thing. If we go to level two, they each took a slightly different approach, but very similar. They each use the lat, they each use text split substitute and large.

[00:12:03.81] Co-host 2: Giles Male: Why was it so? I think they're both good answers. I think the fact the Excel agent has used the, um, array as part of large to pull both the first and the second largest is impressive. So. But yeah, both. Good.

[00:12:17.97] Host: Paul Barnhurst: You know. But what's interesting is over here, try number two. They each used a different answer. So they may have both used lead again. But it wasn't the same formula. Something was different in it either they you know like the first 1st May have done some. The other 1st May have done a plus instead of the array constant. So they changed up the formula a little bit. Yeah I thought that was interesting. Now, if we go to question three very different approaches and not surprising. Rosie took a really complex method. I would never do it this way. It works. But right. Putting in a ray constant for the numbers 1 to 9 and a find when all you need to do is a simple indirect.

[00:12:57.63] Co-host 2: Giles Male: Yeah.

[00:12:58.51] Host: Paul Barnhurst: And then the next time it did the simple indirect. So three of the four times it used the same formula. But this gets back to what we've said before. And we'll continue to see this if I'm getting a different answer every time. What if I'm adding a new year? Am I confident it's going to use the same formula as it did in the prior column, or is it going to try a new way of solving it? Or am I going to be smart enough to prompt it to say, hey, you need to use the same formula as you had for the new year, right? These are little things that I think can confuse people. Exact same answer here. And then, you know, same answer here but different here. And so the more complex it uses the different answers each time. Right. Because you would agree, sorry. This is, uh. You gotta think more through the solution than this one. There's really kind of one way 90% of people do it.

[00:13:44.73] Co-host 2: Giles Male: Yeah, but what? But what? Neither tool has spotted that the answer for level five is one very small tweak to the solution that they had at level four with Len and substitute. So. So they both answered it right. But they've both taken quite an overly complex route to get there.  But it's still right.

[00:14:03.13] Co-host 1: Ian Schnoor : Yeah. I mean, this continues to be very impressive. It's mind boggling actually, that, you know, they can read English instructions, figure out which are the green cells or the colored cells that need to have the answer put in. It understands that and it puts in an answer. Brilliant. At the same time, I continue to believe that any that if you are using this in your work to try to, you know, take an instruction and build a solution, you sure want to make sure that you understand what it has generated as a response, to know if it's reasonable to know what the tool is doing, to know if there's a better way. Um, and so there's a lot of people in the Excel world that don't use the indirect function or the sum product and don't understand what they do. But, you know, the flip side is it's a great learning opportunity to try and take a solution and say, hey, I'm going to learn something by what it's done here.

[00:14:50.40] Host: Paul Barnhurst: Yep. So I think the highlight is great. They all put actual formulas in. Yep. The only place they hard coded. Sorry.

[00:14:58.84] Co-host 2: Giles Male: Bless you.

[00:14:59.68] Host: Paul Barnhurst: Let them edit that out or we'll just bless me I'll.

[00:15:01.96] Co-host 2: Giles Male: Leave. It is. Leave it in. That's great. We're leaving it.

[00:15:04.12] Host: Paul Barnhurst: In. The only place they used text was really around array constants or, you know, setting it to a 1 or 2. So it makes sense. It's not like in some of the others where we've seen a lot of manuals.

[00:15:17.76] Co-host 1: Ian Schnoor : Yeah. None of these, none of these has typed in dead numbers, uh, into which it could do instantly, but a human couldn't do. So they have all generated formulaic answers for the solutions. But fascinating to me, as you said, that they've using, uh, used which? Which makes sense. Two different people would choose two approaches to solve the same problem, and two versions of AI have chosen different ways.

[00:15:43.18] Host: Paul Barnhurst: Each time, though half the time they've chosen something different. Five minutes later, getting the exact same question. That's a little concerning.

[00:15:50.86] Co-host 2: Giles Male: That they are both very impressive sets of solutions. I know they're not perfect, but they've understood.

[00:15:56.98] Co-host 1: Ian Schnoor : I mean, they've got all the solutions here. Yeah, right.

[00:16:00.42] Host: Paul Barnhurst: They've got 1210 out of 1250. That's a great score. The the answers clearly show that they've been trained well on Excel formulas. They're able to reason through complex formulas. And let's just show a quick example. So I think we'll move into testing the agent mode. So this is the same case I just showed you. What we're going to do is I'm not going to do level one and then ask it to do the rest. We're going to have it done all seven.

[00:16:25.14] Co-host 2: Giles Male: Uh, Paul is I'm just conscious, like, we really want to give Ian as much time as possible on this one. Do we even need to show this?

[00:16:31.98] Host: Paul Barnhurst: No. I think we should go to. Let's go to the next one.

[00:16:33.92] Co-host 2: Giles Male: Because you've basically just shown it basically got it right with formulas. Yep.

[00:16:39.28] Host: Paul Barnhurst: I'm with you. All right. So we've been recording that whole time. We're providing our humor. You get to hear us unscripted. We might let him edit it. We'll probably leave it all in. All right, so I asked it open the new cases. They're going to open it. Error installing functions. All right I'm going to pause it for one second. We'll be back with the case up. After some internal discussion, we've decided to skip the Excel athlete case because we just showed you can solve it in the interest of time. If you've watched the other episodes, you've seen it twice. And so where we want to focus on is the humble MVP, the greatest MVP in the world, the musical. It will be coming to Broadway soon. You'll be able to watch it. It's a two hour long production. But in the meantime, until Giles gets the funding to build that, he's had an Excel case built all about it. And so this is our complex Excel case. This is a lot harder than the other one. It has things like the humble closet where we can see Giles has all his fur coats and outfits.

[00:17:43.02] Host: Paul Barnhurst: It has dance formations. So if you're a dancer you might recognize some of these. If you're not, you might be like, what the heck? But they have to figure out different moves and positions and you can see the dancer in the formation. And here's the lyrics to his awesome song. So basically there are the bonus questions and the different levels of the actual case. This is the main stage here. And so what we're going to do is we're going to feed it all this information, and we're going to ask it to solve level one. If it does that right, we're going to solve the other levels. So I put in the same prompt we've used before. We're going to go ahead and hit play. And then I'll give Giles 30s to see if he wants to add any color to the case before we just let it run. So we're going to go ahead and kick off level one. Go ahead. Giles, any color you want to add?

[00:18:32.56] Co-host 2: Giles Male: Only that level one can trip people out. Or we've seen it trip, uh, tools out before where you've got to look at the closets. If you go back to the humble closet tab, you've got to go over to the right hand side. You showed the picture, but what you're actually going to have to do, the tool is going to have to do is look here and it's looking at counting the number of items, um, in each of these, uh, vertical fur coat stacks. So this can be quite tricky. Um, and then it ramps up. So there's other questions about the fur coats, then the dances, spatial logic that you've got to deal with. Uh, so yeah, this is tricky.

[00:19:12.24] Host: Paul Barnhurst: The reclusive recursive. Let's check out some of these names for a second before we pause here. We got the glow stick Arctic rave, the one I'm really curious about. Do you really have the pimp is one of your your shelves in your closet, Giles?

[00:19:26.88] Co-host 2: Giles Male: 100%. Percent. Yeah. Every one of these fur coats I own and handmade. They're great.

[00:19:33.30] Host: Paul Barnhurst: The Queen of Narnia. And that, folks, explains his humbleness. So in the meantime, we'll be back here in a moment. All right, we're back. And we're experiencing what others have shared on LinkedIn and other social media. That's right. Excel agent is in beta. And sometimes what we find is things just hang up forever. And we're experiencing potentially that now it's hard to know. One of our problems with agents is it doesn't tell you where it's at in the process. All you get is, hey, here's this little moving thing at the bottom. I guess it's still working, but you just have no idea how close it is. If it's really working, is it going to stay there forever? So we've made the choice to kill it. Any any thoughts on that?

[00:20:22.62] Co-host 1: Ian Schnoor : Giles, we've been going. How long did you say we've been? We've been going 12 minutes now.

[00:20:26.88] Host: Paul Barnhurst: 11.5, 12 minutes now.

[00:20:28.76] Co-host 1: Ian Schnoor : Believe it or not, that's actually not the same as forever. I know in our day, our current brains, we think that 10s is forever. But it's still to your point, I'm just bugging you. It's a long time, isn't it? And it's a long time. And I agree with you. What I liked about the other tools we looked at is they gave us some sense that they were working, processing, thinking. So this we're not sure. We're just seeing that moving dotted line at the bottom. Maybe it's hung up. Maybe it's working, but it's hard for us to know. Giles.

[00:21:01.68] Co-host 2: Giles Male: Yeah. To add to that, I think the measure that's always in my head is could I have done it faster? And that's been my frustration with ChatGPT and Claude at times. Where you go, oh, God, I could have just done this myself. The first question, once you understand it, albeit there are some tricky areas to it, you could have done that within five minutes. I would have thought if you knew what you were doing.

[00:21:22.00] Host: Paul Barnhurst: You get what, 30? They got 30 minutes for this case.

[00:21:24.54] Co-host 2: Giles Male: 30 minutes for the case. Um, yeah. You. We would have done this quicker. Um, manually.

[00:21:30.26] Host: Paul Barnhurst: Yeah. So level one, you need to be finishing in probably 3 to 5 minutes if you're doing seven levels. And bonus questions.

[00:21:35.38] Co-host 2: Giles Male: Definitely.

[00:21:36.06] Host: Paul Barnhurst: If not less. Ideally, a lot of them are doing 1 to 2 minutes for level one, right. Especially depending on the case. But I mean you're more advanced. Experienced people are 1 to 3 minutes for level one.

[00:21:47.94] Co-host 2: Giles Male: Yeah a maximum.

[00:21:49.54] Host: Paul Barnhurst: Yeah. Yeah. Sometimes 30s if it's a really easy one.

[00:21:52.66] Co-host 1: Ian Schnoor : It's the easiest one in the whole question. Of course isn't it Giles.

[00:21:55.90] Co-host 2: Giles Male: This question is the easiest level I think. Yeah. So we and we asked it to just solve level one. So there you go. We have killed it.

[00:22:02.90] Co-host 1: Ian Schnoor : So we did not perform nearly as well. We didn't perform at all. Um, but it it didn't perform I mean whereas the other tools we looked at did solve this level didn't it. Fairly quickly.

[00:22:13.14] Host: Paul Barnhurst: So, you know, it could be interesting and we'll just give it a minute or two is just say review level one and let me know if you have any questions. Just see how quick it comes back and reads it how it thinks about it. In the interest of time review. Okay. Please review the question. And let.

[00:22:33.32] Co-host 2: Giles Male: Me know if you have any questions.

[00:22:37.40] Host: Paul Barnhurst: And how.

[00:22:38.16] Co-host 2: Giles Male: You would think.

[00:22:39.52] Host: Paul Barnhurst: About solving it. So not asking it to solve it. Just see if it will come back quickly. We'll give it a couple of minutes here and then we'll just kill it if it doesn't give us anything. Yeah. And that's the biggest thing. Agent mode. You know I think the biggest issues we're all experiencing are when we try different things online. One, it's online. It's not on the desktop. Two, you just have no idea. Is it hanging up or is it working. And you know, I think the third thing which I think is a good thing, but you're seeing is one it picks what it's doing. You don't have to select through a bunch of different models, which at times can be a problem. But for I'd say the average user, that's a good thing, right? Not having to try to figure out what model to use.

[00:23:25.35] Co-host 2: Giles Male: I think we kind of set on that after the last episode, which is probably still my position. I think having the tool choose which is the best LM would be great. It doesn't feel like, well, none of them that we've tested so far do that. Having the choice manually sounds great, but actually might add a layer of even more stress and confusion. But I do think the big drawback at the moment, well, they're all drawbacks is being online. Massive drawback. But it's a first stab and I'm sure it will end up on the desktop. Not seeing the thought process I think is a really big drawback at the moment, but again, maybe that will update as well in a future version.

[00:24:04.35] Host: Paul Barnhurst: All right. Well we'll give this about another 30s then we'll just go ahead and kill it and move on.

[00:24:09.55] Co-host 2: Giles Male: Interesting though, isn't it, because all you asked it to do this time was just say, you know, what are you? What are your thoughts on level one? And it seems to be just hanging.

[00:24:18.87] Host: Paul Barnhurst: Part of it could be the day we're testing that now has been out a week or a ton more people on it. And, you know, the bandwidth isn't quite there yet. Maybe hard to know, but we definitely, you know, people have complained about it hanging and we're now experiencing it. This is the first time I've really had this particular experience. All right. We're going to kill it. We're going to close this and we'll be back in a minute. With the FP&A use case. We'll run through a couple of those and see if we can get them to work. With any luck, you won't be listening to us. Just ramble with nothing happening on the screen. So we'll be back here in a moment. All right. We're back. And now we're going to cover some FP&A use cases. We're going to start the first one with, uh, what we call a PVM. Price volume mix could be done on gross margin. It can be done on price on a number of different things. This is a case we do with our FP&A students and our courses. It's called Paul's Getting Spicy. It's an inside joke because every so often I get a little grumpy and my Copartner says depending on how many peppers he calls it. I tell him I'm deciding whether he'll come to the meeting or not.

[00:25:24.87] Host: Paul Barnhurst: So he labeled that to have fun. But lets get to the real point. Basically, it's a smoothie business. There's three types of smoothies. There's BB, Barnhurst Basics, PMP, Polly's Mango Punch, and the Boss Barnhurst Organic Superfood Smoothie. So we'll see. Here we have units or volume of what was sold. You can see the percentage. So this is what we budgeted. You can see BB was supposed to be 25%. These others 38. We can see what the gross margin was supposed to be, the gross margin per unit, how much margin we get on each unit, the overall. Then we can see the actuals. In actuality, we sold more units, different mix, different gross margin per unit. So what we're going to ask the tool to do and we can see we have the formulas in here. So this is fairly straightforward. We're going to ask it to calculate the price. How much was due to volume in total, and then how much was driven by mix for each product? How much was the mix of the volume for each individual project so we can understand the mix? It should come back with an answer of 220. For the first one, we're going to ask it to just solve the price.

[00:26:34.53] Host: Paul Barnhurst: Then we'll ask it to solve the next two. So this should be a fairly quick return. But let's go ahead and kick it off and see how long it takes. All right. We'll be back in a moment. After several minutes of discussing things like famous parents and singing and videos and all kinds of other stuff, as we waited, we decided to come back in. And you can see Agent Mode is still spinning. What we don't know is, is this due to agents in general not working? Is this something with my account? But these are the type of things you're going to deal with in real life, especially if you're trying to use a tool in beta. And so it's just something you have to be aware of. You know, Excel labs has worked great many times for me, so I'm still a proponent of agent mode. I think it can do great things, but at the moment it's not working well and that's always a challenge of live testing. So what we're going to do, we're going to get in set up and we're going to do his testing here in a minute. But before we do that in Giles, anything you want to add to this section of our testing? Anything you want to tell the the audience?

[00:27:37.67] Co-host 1: Ian Schnoor : I'm still reeling from the news that Giles comes from a famous progeny. I'm. But you know what? Everything makes so much more sense now. It all comes together. The white fur coat, the, uh, the singing and dancing, the performance art. I it's like it clicked for me. Paul, what about you?

[00:27:57.35] Host: Paul Barnhurst: It did knowing he has, you know, parents that are quite musical and in the music that makes sense of his videos and his productions. And we could see where he gets that, uh, incredible talent from.

[00:28:06.99] Co-host 2: Giles Male: Well, I have none of it. I just know, luckily, my sister and my brother in law are very talented, so I'm just using their talent.

[00:28:14.57] Co-host 1: Ian Schnoor : Yeah. Well, I, I don't agree with that part, but it all hangs together now. As for the tool, yeah. I mean listen you're right, Paul, maybe this is just your account or your system acting up. However, this is what people potentially experience. Let's see if I have any better luck with um, I again, I said this before. I wish, though, unlike the other tools, that the agent was giving me some sense that it was processing, thinking, working, refining an answer. It gets frustrating when you just see it, the dots and bars moving without any sense for whether you're hung up or whether it's going to generate an answer.

[00:28:51.05] Host: Paul Barnhurst: And it's clearly not working. It's not just hung up. I just ask it, you there? And it hasn't responded yet.

[00:28:59.21] Co-host 1: Ian Schnoor : It's deep. It's deep in thought, deep in thought.

[00:29:02.69] Host: Paul Barnhurst: That should not take long to respond to.

[00:29:06.49] Co-host 2: Giles Male: Well, can I just add one thing? Yes. And again it's just an observation because I've actually seen, uh, some posts from Microsoft people sometimes resharing other things from other people. The, the difference between the message you get or the assumption you might make about the its performance from what you read versus this, we don't know what the cause is, is night and day and that and again, that is the thing that gets me triggered a little bit is like, God, we just need to be objective and honest about this stuff. It's not perfect, it's beta. And I don't know why. The messaging isn't just that. Try it. Let's help it get better.

[00:29:47.39] Host: Paul Barnhurst: Alrighty, I'm going to pause. We're going to switch to end screen and we'll see how this works. Back in a moment. We are back with Excel agent mode set up on another person's computer.

[00:30:01.51] Co-host 1: Ian Schnoor : It didn't like you, Paul didn't like you. Let's see if it likes me any better. I mean, this is now my screen. You're seeing the you're seeing in the cloud and not not my favorite environment, but for for spreadsheets. But in the in the in the cloud, we've got the Henderson AFM model, the model that you know people a full three statement model of a company with a cover, an executive summary, an assumption sheet, the scenarios, and then the actual model. This is exactly this classic three statement model. It's exactly what people need to build on the AFM exam. And so I am going to get to see if an AI agent in agent mode can answer some questions. Let's start with the simple one I've been asking. Are there any hidden sheets in this file? And I sure hope my version works better than yours. Um, you can see though, uh, just like. Oh, here though. Look, um.

[00:30:57.53] Host: Paul Barnhurst: Within a few. Yeah, I like you. I guess Microsoft knows who to treat.

[00:31:02.09] Co-host 1: Ian Schnoor : Well, it was verified. And if I recall, I think I'm. So Giles can't even get the software. Uh, if I recall, yours is hanging up on you. I might be the only Non-microsoft MVP in this room. And mine is working, isn't it? What do you make of that?

[00:31:18.67] Co-host 2: Giles Male: Yeah. Well, I'm being thrifty because I should be able to download copilot for free as an MVP, and I couldn't work out how to do it, so I haven't done it yet.

[00:31:27.99] Host: Paul Barnhurst: Giles and I are just being blackballed by Microsoft. No, uh, no.

[00:31:32.71] Co-host 2: Giles Male: We say too much. We're too vocal on LinkedIn. We're being.

[00:31:36.51] Co-host 1: Ian Schnoor : I don't.

[00:31:36.79] Host: Paul Barnhurst: Know our status. We love you, Microsoft.

[00:31:39.83] Co-host 1: Ian Schnoor : But look at that. It, um. It did work, right? So there are hidden sheets. It knows that the covert sheet is hidden. The clandestine sheet is very hidden. And it cannot be unhidden by the Excel user interface. Very nice. So. And it even tells us how it. So it says there were two hidden sheets. One hidden, one very hidden. How I checked it. Queried each worksheet's visibility and listed any that aren't visible. Okay. Um, nice. Well, it passed that test. What about again? Are there any white values on this sheet? Um, we discovered this. Sometimes finding cells that were painted white worked. And sometimes it did not work. Uh, with the agents we were looking at. So let's see if this does it. Now, this one is not. Let me just double check. They should be here. I don't know where they are, but this file should have some. There we go. Yeah. So I do have some hidden sheets here. We can see in my formula bar that there are some, uh, there are some cells that have been painted white. And let's see if, um, if it can it's, it's reasoning now. It tells it sort of stopped for a second and it's checking and it's working.

[00:32:52.77] Co-host 1: Ian Schnoor : It says checking for hidden values. Although I looked, I asked it to look for white values. And it's sort of understood that white means hidden. So it's doing that. But it's um. Again, we don't have a good sense for how close it is. Again, Giles, to your point, to your great point earlier, again, I do the same thing. I like to say, could I have done this myself faster and and, um, and again, control a control and the letter A selects all so that, you know, doing this manually takes no more than 10s control A selects the whole sheet. And then going home and then to the, you know, shading color. If I was to shade the entire background gray, I can instantly see. And all I have to do is scroll down, I can, I can zoom down, I can scroll down and I will see them so clearly much, much faster to do this right now on my own. But it's exploring different approaches. Well, we might just stop this one because it's not giving me much guidance, is it? Oh wow. Okay. This one, it took 70 seconds. This one says no white font values were found.

[00:34:04.30] Host: Paul Barnhurst: It tells you what range it checked. Those fonts were right outside your range, weren't they? That range was checked, which is really weird. Weren't they like it or something?

[00:34:11.86] Co-host 1: Ian Schnoor : No. They're inside the range, I believe. Let's see. They are.

[00:34:15.74] Host: Paul Barnhurst: They are okay. I just.

[00:34:17.14] Co-host 1: Ian Schnoor : S.

[00:34:18.06] Host: Paul Barnhurst: Okay.

[00:34:18.98] Co-host 1: Ian Schnoor : Yeah. They're inside the range. They're here, they're S2 and it says it checked B1 to you. 340. So that's fascinating because um and again I'm using this. So to me we use these sorts of tests just as a general as a general check on, you know, capabilities. Right. So we're able to do that. Now let me see again if I can find any inconsistencies. We know that this formula has a dead hard coded value at the end, which is a bad thing in the world of modeling, as does this cell here. Let's just see if I can ask for it to look for. Are there any inconsistent formulas on this sheet? And, you know, it might take a few minutes, but we. But we know that it's working, right. We know that it's, um. Here we go.

[00:35:04.56] Host: Paul Barnhurst: So we know that it likes you. So that's good.

[00:35:06.84] Co-host 1: Ian Schnoor : I'm a little bit concerned, you know, back to yours. Uh, you know, I'm a little wondering whether there's something up with your. I mean, we can't know, but at least on mine, it's giving me some minor prompting. Right? And it's changing this. So. And I'm also noticing mine, I'm.

[00:35:23.44] Host: Paul Barnhurst: Also using edge. I'm going to try Google Chrome and see if it works.

[00:35:27.72] Co-host 1: Ian Schnoor : You'd think that it would have a more successful time with the browser I know.

[00:35:32.80] Host: Paul Barnhurst: That's going to be my ironic moment here.

[00:35:35.60] Co-host 2: Giles Male: What? A bit like a bit like you would think using teams would be easier than using zooms. If you're a Microsoft person.

[00:35:41.64] Co-host 1: Ian Schnoor : You might think that too, but mine. Paul, I don't know if this mine is changing colors. The little dancing dotted line is changing colors. I don't think I saw that changing colors on yours was it or am I?

[00:35:54.20] Host: Paul Barnhurst: I don't think it was.

[00:35:55.68] Co-host 1: Ian Schnoor : I thought it was just read the whole time. But anyway, it's.

[00:35:58.44] Co-host 2: Giles Male: Just the fact you've got text coming up shows that your maybe pool wasn't quite working the same way, but, um.

[00:36:04.86] Co-host 1: Ian Schnoor : Right, right. Let's see. Yes. Inconsistent formulas were found well, totaling inconsistent 392. Um, that's a lot. That's a lot. But see here we know that. Row 13. I'm just gonna, I mean, I'm not sure. Sometimes it's giving me some of these could be hallucinations and and and actually correct. But the one I'm looking for is row 13, which has a hard coded value. And, uh, I don't see. Do you guys see? I don't see anything that's pointing out to row 13 that it has the error seeing anything.

[00:36:41.78] Co-host 2: Giles Male: So it's giving you R1, C1 references, which is a bit, uh.

[00:36:47.94] Co-host 1: Ian Schnoor : Confusing, a little unusual. And for those of you who are not familiar with that, there are two ways to, to denote cell references, In fact, would you? I think 99.9999999% of Excel users use, um, I don't know. I forget what the name of it is, but the cell name is D2. But you can also call it, you know, you can use the RC. I don't even know. I don't even know why. Maybe it's from from a programming standpoint, perhaps.

[00:37:18.32] Host: Paul Barnhurst: I'm sure it has to be from a programmer or that.

[00:37:20.64] Co-host 1: Ian Schnoor : Must be what it is.

[00:37:21.36] Host: Paul Barnhurst: Right?

[00:37:22.52] Co-host 2: Giles Male: I think it is. I've used it a little bit in esports for map cases, but beyond that I've never used it in modeling, to be honest.

[00:37:29.48] Co-host 1: Ian Schnoor : Well, let's see one here. H109. Yeah, I just mean, in typical grid work I have not seen people tend to use it. J that's just your J 28 I don't know what it's uh, J 28 so row 28 is just the gross minus the freight. This formula is the same all the way across, and it tells me it was expecting. Uh, I'm not even sure how to read this, to be honest. I'm not sure what it means when I was expecting our row 57. I don't know why it was looking to row 57. But that doesn't make any sense to me, so I'm not sure. Giles, do you have a sense for what this is?

[00:38:07.42] Co-host 2: Giles Male: No. Other than obviously the numbers are going to be a reference from the starting cell, so it kind of makes it even more difficult. So it's actually from. Yeah. Okay. What's that next bit though. Is that more useful? The hot spots go up a little bit.

[00:38:23.30] Co-host 1: Ian Schnoor : Hot spot row ranges with multiple deviations. Rows 108 to 110. The ratio is 142. But the problem is even in the first section here we're not understanding what it's.

[00:38:34.38] Host: Paul Barnhurst: Why don't you just ask it to give you, uh, traditional references instead of our, you know, cr r1c1 notation and see if it then we can at least try to understand it, because how often we use r1c1, like you said in your modeling, never.

[00:38:54.06] Co-host 1: Ian Schnoor : Uh, can you please use traditional references as opposed to r1c1 notation? Sure. Um, let's see if it can do that.

[00:39:07.12] Co-host 2: Giles Male: I had to Google it because I wasn't sure. It's A1 notation. So interestingly, you're not now getting the text on the side. The color is still changing. Uh, but you're not, you're not getting that.

[00:39:19.44] Host: Paul Barnhurst: I open mine in Google Chrome and it's still just frozen. I even just said hello. Why do you keep hanging up? It won't even answer that question.

[00:39:29.08] Co-host 2: Giles Male: And you might have to scroll down. I think there's more that's been written underneath that.

[00:39:33.16] Co-host 1: Ian Schnoor : Yes, yes, I can use the traditional A1 style going forward. I'll show formulas like D1. Okay. So anyway, I think we can agree though, that, uh, it was not. I mean, even in spite of that, it was not um, it.

[00:39:43.84] Host: Paul Barnhurst: Was not finding what we wanted.

[00:39:45.28] Co-host 1: Ian Schnoor : Impressive. No. So what I'm going to ask you to do, why don't we. It did not find it. Um. That was. I thought that was going to be a little bit different. Let's try. Why don't we try this? Why don't we try? I'm not even going to ask it. I'm not even going to ask it to tell me why the balance sheet wasn't balanced. I'm not feeling optimistic that it will determine that. But what I am going to do is ask it to build a model. We're probably going to need to pause for this, because when I tried this earlier, it did do something and I was pleasantly surprised, but I'm going to ask it to build a five year forecast model on the model.

[00:40:17.50] Co-host 2: Giles Male: Where are you doing this, by the way? Because I can't, uh, we can't see anything moving on the screen.

[00:40:22.94] Co-host 1: Ian Schnoor : You do see it. Okay. So this is the starting point of the model. This is just the three years of historical financial statements for Henderson. This is the starting point of any financial model. Start with the historical financial statements. This is the starting point of the model exam, the AFM exam. So this is how people start their models: a historical set of financial statements and we build from there. I have simply said, build a five year forecast model on the model sheet for the years 2025 2029. Build separate schedules for revenues, cost, depreciation, income tax, working capital, debt and equity. And I've said make reasonable assumptions and I want to see what it can do. So I suspect that we are going to this is probably going to take a few minutes. But let's just see if we start to get a prompt here and then um, and you're we're timing this one so we can see how long this is going to take. But it already said it's planning. It's organized when it does give you feedback. I'm kind of enjoying this feedback. It's not too much. But it's first like at least I know it's thinking, right. It was planned. It's doing something here. It's organized. And now it's defining. It's planning, organizing, defining. So it's doing something. Uh, so I like that.

[00:41:36.52] Co-host 2: Giles Male: We've seen a few versions of this as well haven't we. So we've seen one where we got quite a lot of information shown and it stayed there. And you had to scroll back up, I think. Was it Rosie last week where it showed it but then took it away? It was one of the two, I think that's right.

[00:41:49.40] Co-host 1: Ian Schnoor : That's right. It would show it and then.

[00:41:51.34] Co-host 2: Giles Male: And then it would go again. Yeah, I find that I find that frustrating because I wanted to read, even if it was getting it wrong. I wanted to read it and see what was wrong.

[00:42:01.18] Co-host 1: Ian Schnoor : That's right. And we'll see if of the two tools we looked at previously, one of them gave us solutions and then said, do you want to apply them? And you had to click apply, right? Whereas the other one, you, it just went right in and did it. So this is, this one is actually saying finally it's making it seems to be making some good progress here. This is now finalizing the schedule set up. But um, we can see it's been running for only about a minute or so. Right, Paul I think uh, so far and it's implemented. It's got a lot of good action words here. Right. It was defining. It was, now it's finalizing and implementing. It's doing something. I haven't seen it. I haven't seen it. Put anything into my spreadsheet yet, I don't. Right. Um, you know what I've discovered? Sometimes it adds things, and you're not even aware it's adding things down below, but it, outside of the sight range. But it is working here obviously. Uh.

[00:42:54.65] Co-host 2: Giles Male: There's a lot for it to choose from as well, you know, because if you just think about the prompt, you said build a five year forecast from 2025 to 2029. So it has to interpret that you've got three years of data in there. It's got to find the data in row seven. I think it is to see that those three years are the prior years. If you think of what a modeler might do, you could go truly vertical on the same worksheet, like a lot of AFM people do. You could go my sort of fast approach where you pull all of it out, you have a separate inputs tab. You don't do any of the outputs on the same starting model tab, like so how is it going to make those decisions? Um, if it did either one, uh, or something in the middle, I'd be pretty impressed.

[00:43:38.97] Co-host 1: Ian Schnoor : Agreed, agreed. Now we, uh, formatting and final checks. Checks. Now we, you know, the other tools did attempt something not quite ready for prime time, but let's see, what.

[00:43:52.39] Host: Paul Barnhurst: Should we pause and come back when it's done?

[00:43:54.59] Co-host 1: Ian Schnoor : Or why don't we pause? I again think it's helpful for people to see, though, that there is definitely some waiting. But why don't we pause and then we'll let you know how long it took once we come back and we'll show you exactly what it looks like as soon as we have something on the screen and it has stopped working.

[00:44:08.63] Host: Paul Barnhurst: We are back. It's insane. No. In all seriousness, we've been very impressed with the last, uh, round of what we got built. But enlighten. Take us through a little bit. The reason for what, 11.5 minutes? And what do we get?

[00:44:22.23] Co-host 1: Ian Schnoor : Yeah, yeah. Amazing. Just under 12 minutes. It, uh, it iterated, but it showed me what it was doing, and it took about 9 or 10 minutes until it started to start populating my, uh, my spreadsheet. I always like watching it because it often makes mistakes at the beginning, but then corrects them and fixes them. So listen, this is pretty, you know, pretty impressive and extremely impressive. I have a forecast of the income statement. Just eyeballing it, I can see that the revenue directionally is moving in the right direction. We've got costs. It is building links. It is just linking the revenues and costs down below to schedules. I asked it to build schedules. Um, it has got something. It's got nothing for cost adjustments, but it's going down to EBITDA. It has a link to a depreciation schedule, a link to a schedule, and it's got taxes. Now I am wondering I'm not sure what it did on taxes, but it has assumed all taxes are going to be cash and no deferred taxes. So that's the kind of thing I need to understand as a modeler to know if I'm comfortable. Why did it do that or not? And the total here is also coming from below and net income. So it's not a great job building that. And then the cash flow statement is added back.

[00:45:31.65] Co-host 1: Ian Schnoor : It's adding back the depreciation from the depreciation schedule. So it's done some impressive work. It's pulling in changes in working capital. It made some assumptions on the capital expenditures here. I'm not sure we'll go see where that came from, but the point is it has a fully integrated, complete income statement and balance sheet. Look at the cash flow statement. Let's look at the balance sheet. The balance sheet is built. And just eyeballing it it's actually balanced right. Look at every year that it has a five year forecast balance sheet. It's actually balanced, eyeballing it now a financial model. A good financial modeler would know that you don't ever want to see negative cash on a balance sheet. So we'll see what we can do about that. And I think it's giving me a warning about that down below actually. So look what it says it did. It actually says it built a five year forecast. It said it added five years. It tells me all the work that it did here. It built a driver's based model rolling forward gross revenue by year over year over growth rate. A CapEx is a percentage of net revenue depreciation. So it tells us how it calculated everything. It tells us the assumptions it used, which are.

[00:46:38.23] Host: Paul Barnhurst: What I like. It also says the assumptions are editable. Like hey, go ahead and modify them. We actually built them properly.

[00:46:44.21] Co-host 1: Ian Schnoor : And then it even says, you look at this ending cash trends negative because we modeled no revolver draws. That's like, if you like, I can add an automatic revolver to keep cash greater than or equal to zero for a minimum cash target. That's impressive. And wanted a sensitivity table. I am going to ask it to add a revolver to see what it does on that front. The working capital looks reasonable here. The net PPE looks reasonable. I have to see what it did. It did not use a revolver. The deferred income tax is a curiosity. It has just linked it to one single cell. So it keeps the deferred taxes constant up above.

[00:47:20.37] Host: Paul Barnhurst: Didn't it show no deferred income taxes?

[00:47:23.01] Co-host 1: Ian Schnoor : No deferred income tax expense, which would be if you had no deferred tax expense.

[00:47:27.13] Host: Paul Barnhurst: That you tell it to the last thing you're never.

[00:47:29.13] Co-host 1: Ian Schnoor : Yeah, that would keep the balance sheet constant. But I'd still prefer to build in a formula. But it's okay. It's doing this. It's got long term debt. So this is quite impressive here. It actually made some very reasonable assumptions about revenue growth. It made some assumptions about freight and warehousing, cost of goods sold, tax rate, etc. receivable days. It made some. It made working capital days assumptions and it made some assumptions around long term debt and an annual debt repayment. Right. Impressive. And it has the little schedules down below. So this is pretty pretty wild. It's assuming a growth rate for revenues and cost. It is built in a fixed asset schedule.

[00:48:09.39] Host: Paul Barnhurst: What did it cost? Because I noticed the Cogs went down but the revenue went up. So did it. So the cogs are lower than, uh, it's.

[00:48:18.31] Co-host 1: Ian Schnoor : Assuming that costs are just.

[00:48:20.39] Host: Paul Barnhurst: A.

[00:48:21.19] Co-host 1: Ian Schnoor : Percentage of revenue. So not my favorite way, but I, I told it to just make some.

[00:48:25.79] Host: Paul Barnhurst: Yeah, yeah, yeah. I mean acceptable given what we told you.

[00:48:28.63] Co-host 1: Ian Schnoor : Totally. Totally. It's not you know, it's not, it's not factoring, um, variable and fixed costs. But that's okay. For this purpose, it built working capital. I'm really impressed with this one. It calculated the change in working capital. Um, which is correct. It's the actual change in the working capital. This is the working capital, the assets minus liabilities. It did that right. The change in working capital is the current year minus the prior year. However, for a financial model, we're usually talking about the change in cash, which needs to be the other way. So the sign should be negative. But I'm really impressed because it actually knew how to do that. So on the cash flow statement, it flipped the sign and it made that negative. So it is the correct movement of cash here. This is pretty good. I mean this is better than good. This is incredible.

[00:49:11.93] Co-host 2: Giles Male: How did it actually work out the balances for all the working capital lines as well. What did it do? Can you see.

[00:49:18.25] Host: Paul Barnhurst: So it made the revenue assumptions.

[00:49:21.05] Co-host 1: Ian Schnoor : It's taken the revenues. And then it is calculating the receivable balances as a revenue time. It's a day's assumption. Yeah. It assumed 47 days receivable. And I'm going to guess it figured out that that's what the historical days were, but I'm not sure. Uh, it must be because the working capital balances are, uh, are reasonable, right? They are growing at a reasonable pace in the future years. They all look reasonable, the prepaid expenses look reasonable, and as.

[00:49:53.23] Host: Paul Barnhurst: Does use a historical or pretty close to an historical average.

[00:49:56.55] Co-host 1: Ian Schnoor : Yeah. So it's using a very reasonable assumption. It made a debt assumption here. So let's go a step further.

[00:50:02.63] Host: Paul Barnhurst: So far this is the best we've seen yet in all the time we've run.

[00:50:06.23] Co-host 1: Ian Schnoor : So let's go to the bottom and say if you'd like I can add an automatic revolver. So now I'm going to push this a little bit harder. And please add an automatic revolver to keep cash. Right. I'm going to say greater than or greater than or equal to. Right. Um greater than or equal to zero. And I'll say borrow whenever there is a cash shortfall. Fall and repay the revolver whenever the company has excess cash. Right. Give it some instruction. Let's see what happens then. If we ask it to build a revolver again. I am still with you guys, I believe. I mean, the ability to still do this, right? The ability to ask and to challenge and to get it to make, you know, requires that we have an understanding of how this works. I would never want to just hand this in like this, because I think there's already an issue on deferred taxes, but this is an incredible start. I'm not going to lie.

[00:51:07.73] Co-host 2: Giles Male: Well, and also, I mean, there's two things I would say. Number one, I think it's something we've talked about a few times. What does this mean for the other tools? If Microsoft's agent is that many steps ahead of where they're at? The other thing I would say is actually offer a comment, um, Hillary Smart from opera put online the other day where she was saying, you know, there's a very big difference between getting a model mechanically sound, of course, and then right, for a project. I'm really impressed by this potential for it to get close to a mechanically sound model. But then there's a huge leap between that and what are the exact situations and the assumptions you have to manage on your project? Should interest be calculated on like a three over 360 basis or, you know, whatever actuals over 365. So those are two different things. But this is really impressive, especially compared to everything else we've seen.

[00:52:04.00] Host: Paul Barnhurst: You can also tell it to do those type of things right. You can say hey now use 360.

[00:52:08.84] Co-host 2: Giles Male: I mean you probably could, but you but you'd have to. This is my point about you have to then have modeling skill to know the next level. Because if you have no modeling skill and you go, hey, build me a financial model that does X, you need to know the next questions and prompts. But I'm still blown away. I think it's really good.

[00:52:26.36] Host: Paul Barnhurst: Yeah. This is better than any I expected at this point.

[00:52:29.08] Co-host 1: Ian Schnoor : Well, and this is why this is why a lot of financial modelers feel that there's always going to be a part of modeling that that requires. I mean, I you know, I can pretty much guarantee you that for a long time coming that when you are ready to present your analysis on your infrastructure project or your investment opportunity or your credit assessment, um, your client, your credit committee, your boss is going to look at you in the eye and ask you, and you are going to need to be able to explain and justify and talk about how you derive the assumptions and great use, uh, you know, use your AI tool to help you with that. And they're going to still want to make sure you understand every element of it. I completely agree with you. However, this is still a very impressive mechanical effort, isn't it, in terms of its ability to do this work very quickly. I think we can agree this it did this faster in about 12 minutes.

[00:53:22.94] Host: Paul Barnhurst: What if you got this to start with some tweaking, you could probably give it all the instructions. It could get pretty close on an AFM if it continued to be this consistent.

[00:53:34.60] Co-host 1: Ian Schnoor : What we haven't tried, what we don't know is its ability.

[00:53:37.64] Host: Paul Barnhurst: So scenarios.

[00:53:39.56] Co-host 1: Ian Schnoor : I asked it to build this by just making some reasonable assumptions, and it has made some very basic, reasonable assumptions around growth rates on revenue. We would all suggest that a good model should be drivers based and that revenue should factor in prices and volumes. But I assume if I gave it some price volume information, it would figure it out, but I would test it for that. What we don't know is if we gave it a 2 or 3 page case study and we can try that, you know, next, if we gave it a 2 or 3 page case study that simulates reality, could it then factor all that information in? There's a lot more information. Perhaps it could. We will try that at some point here. Now it is um, it has made just some very simple assumptions around margins. It's assumed effectively it has assumed the margin. It's assumed the cost of goods sold as a percentage of revenue. Now it's still working on the revolver issue. What you know, we talked about during our when we were offline is will that enable will that lead to a situation of circularity in this model when when it builds in the revolver we will see. Time will tell. So we've been going here. How long has the revolver question been running? A good five minutes I think about this.

[00:54:48.78] Host: Paul Barnhurst: Point probably about five minutes.

[00:54:50.18] Co-host 1: Ian Schnoor : Why don't we take a pause and come back when the revolver is built and see if it is still if everything is still working well, and then we'll show everyone that.

[00:54:57.62] Host: Paul Barnhurst: Welcome back. We showed you last time it built the model. We've now asked it to do the revolver. We're going to turn it over to show that I unfortunately have to step away for a minute. So Giles is going to close this episode. I'm trusting you with the mic, Giles. Be humble and bring it home. Okay? All right. In the meantime, go ahead and share what happened here.

[00:55:19.62] Co-host 1: Ian Schnoor : So we asked them to build an automatic revolver. We said borrow whenever there's a shortfall and repay the revolver whenever the company has excess cash. And so what's happening here is I don't see it on the debt schedule. The debt schedule is still just showing the term debt, the long term debt. However, when I scroll up, we will see. And it also tells us what it did it it went through and it put it in, um, it set the cat revolver, made sure that that there was cash never went below zero. If the pre revolver ending cash was less than zero, it drew. So that's exactly what it's supposed to do. If the pre revolving cash was less than zero it would borrow. And if the pre revolving ending cash was greater than zero it would repay it to the to the revolver. So very impressive. That's exactly what we want to do. And then it linked it up. And then it verified that the ending cash is negative. Negative. So very very impressive. Now it says um minimum cash target was zero. No interest though there was no interest built in and no hard cap on the revolver. Availability. Great. I mean, it's very, very impressive what it did.

[00:56:28.30] Co-host 1: Ian Schnoor : Let's see here. The balance sheet is still balanced. We can see here the balance sheet is still balanced. But now we actually have a bank revolver and a revolving credit facility in the model. And now it's a formula. It's taking the prior year plus the change it's doing that uh, and you can see cash is now zero. It's not negative anymore. It's borrowing but it's taking it's taking the prior year's revolver and it's adding in the change from the financing activities. Now this is a bit of a long formula. And if you follow through it, it's calculating the change in cash flow, the change in cash, the change in cash flow each year. And it's saying if that change in cash flow is less than zero, then borrow. Otherwise pay off. Uh, not the most elegant. And you know, as a modeler, we could give it guidance to improve it, to fix it, but it's still getting the right answer, so I don't know. Giles, I would say I'm quite impressed with this. I still wouldn't want to have it run in 15 minutes. Hand it to someone. But impressive. It can get this far.

[00:57:34.64] Co-host 2: Giles Male: I'm really impressed. And actually, you know, it's I think, I mean, we would have to test it further mechanically. It's doing a lot of things that the other tools were nowhere near doing, which is fantastic. I actually posted about this today. This is the exact point I was warning about on LinkedIn today though, which is we're going to go through a phase now where it's going to look exciting and you go, oh my God, the balance sheet balances. It works. And I feel like this is the most dangerous period where you go, hang on a minute, because I think you were saying even when you tested this a couple of days ago, actually, it didn't get the right answer. And there were 1 or 2 mistakes. That's right. The thought of going from this, which is immensely impressive, to oh my God, I can just click a button and I know it's going to be right. Balance sheet balances handed in. That's so dangerous. However, a massive step in the right direction.

[00:58:29.62] Co-host 1: Ian Schnoor : Yeah. No. And again, as a good partner, as a partner to help you with your work and to get started and give you a boost for sure. But, um, I need to be able to challenge this new partner of mine and double check. And again, I still have on the balance sheet my deferred income taxes, which is just dead. It's just a straight link. It's a straight up link to a cell. It just said assume the deferred income tax was constant at $8 million per year. I have no idea why it said to assume it was constant. I guess it made a very simplifying assumption for income tax. Again, we could ask it to run a couple more tests on income tax, but suffice it to say, this far, uh, maybe we'll do some more at a later time. Giles. But this far it has. Uh, it's impressed us, and I think tread carefully and tread cautiously and understand what it's doing. But, you know, in a half an hour here generated with a couple revisions, pretty robust. Helpful tool so far.

[00:59:29.24] Co-host 2: Giles Male: Yeah. I think this is, um, I mean, bear in mind this is beta mode. There have been some wild claims from third party tools, and obviously we will get around to all of them. But weirdly, it kind of struggled with some of the easier audit kind of review things that you asked. Yeah, I would have thought it would have found easier. And then this, which is right. This requires far more kind of accounting and modeling knowledge. It seems to have done a much better job. So I don't really understand that. Well, why don't we.

[00:59:59.96] Co-host 1: Ian Schnoor : Try one last one? You're right. It was struggling on the easier. And why is the balance sheet not balanced? Because in this model I deliberately unbalanced the balance sheet. And you can see here that it's out of balance. So I'm going to say why is the balance sheet not balanced. Let's see if it is clever enough to pick that up.

[01:00:18.62] Co-host 2: Giles Male: While it's doing that. Just off the top of my head, we're probably of the same mindset. I'd love to spend more time on this and give it more detail as a prompt. Because you know what? If you did say like, okay, well, you know, there's debt sculpting that you have to deal with or like reducing balance depreciation. I get the sense that there may have been a chance that if we'd given it the specific prompts to do that, like you did with the revolver, maybe it would have done them. I don't know, but you have to know what you're looking for and know what to ask.

[01:00:47.58] Co-host 1: Ian Schnoor : Oh, absolutely. You need to know what to ask. You need to know what you're looking for, and you need to know how to construct it, at least in your mind, so you can test, challenge, and make sure it's quite reasonable that it will come back with lots of very long formulas. And you'll need to know what they're doing and how they are all working to deliver them to your team. So absolutely. And we'll see here if it gives us an answer to my question, why is the balance sheet not balanced in this model. And it's saying analyzing financial and why is it balanced? It's changed. Why is it balanced? Not balanced. It's investing in range issues. It's searching for the balance sheet. It's trying, I mean again, impressive. It can see it and be aware of it. We're asking it to do a lot here. But did you want to just keep watching or pause and then come back.

[01:01:29.81] Co-host 2: Giles Male: I think there's pause. We'll see how it gets on. We'll give it a few minutes and then we'll come back. All right. And we're back. How did it do?

[01:01:35.81] Co-host 1: Ian Schnoor : Yeah I mean it was really interesting on this one. It worked for about three minutes 162 seconds. And again, I asked why is this balance sheet not balanced? I think we're not thrilled about its tendency, as we know with these tools to make things up. It says the root cause in A1. I'm not sure why it's sending. It's A1, it's off by one column references in the balance sheet totals. So totals don't point to the same column components. So it says k 167. This one here k 167. It's currently L 161 to plus L 165. And it should be K. Well actually it is k okay. Like this isn't true. It's making that up. The total assets is currently L 161 plus L 165. That's not true. That K 167 is the correct answer of K 161 plus k 165. And it says k 180 6K1 86. This one here total liabilities plus equity it's saying is currently L 184. So it's making that up. It's not. This cell is not linking to L. It's linking to it's linking to K. So I'm not sure. It doesn't seem to know. And I'm looking at it here and I did not realize or figure out that we have a working capital change in working capital sign issue which is what's causing it off. So definitely not solving it perfectly by any stretch.

[01:02:57.47] Co-host 2: Giles Male: So bizarre they to again that I don't understand the underlying logic because if it can pretty successfully build, you know, a simple mechanical model, you would have thought it would also be able to review and find errors where what you've got in front of you doesn't match that logic. So I don't know. But I mean, if we were to kind of wrap this up for me, I'll go first, then I'll hand it over to you. Ian, quickly. So Agent Mode has not done so well in some tasks, but then actually in the financial modeling task, it's done really well mechanically. It's done some things that we would not want to be done in a certain way as professional modelers, but that is still very impressive. And for me, still now entering a phase where that's very dangerous to people that don't know what they have to do as the next step.

[01:03:46.73] Co-host 1: Ian Schnoor : Yeah. I mean, listen. Agreed. Uh, really exciting to see that. Yeah. I don't understand why it did better building. It did better when I asked it to build a model from scratch than it did checking on existing models. But that means you got to be careful. It was doing. It seemed like it was doing some hallucinations in its answers when it was looking at an existing model. Again, extremely impressed that it could take a historical set of statements and build up a forecast. But you're right, there will be some line items that it seems like you are not correct, that you don't like, and you'll need to be able to provide guidance and ask why there's a mistake and how to fix the mistake, etc.. So, but overall this is just beta and version one. It's exciting to see and think about what's going to happen in the future. However, you know, as we've been saying, it does not mitigate the requirement to fully understand every every element of it. If you truly want to have credibility and impress your colleagues and your team and be reviewed as someone who has a great level of understanding so it doesn't replace that, I don't think. But it's pretty wild that it can do this very quickly. So that's kind of what I've got. Maybe we'll do some more testing on this tool next time, because I think there's some other things we could do.

[01:04:52.07] Co-host 2: Giles Male: All right. Well there you go. That wraps up this episode. Thanks, Ian. Thanks. Paul. Who's, uh, ducked off early, and, uh, we will see you all in the next one.

[01:05:00.87] Co-host 1: Ian Schnoor : See you in the next one. Take care.

[01:05:03.75] Music: The mod squad. We are the Mod squad.