Excel Hour - Learn to Excel with The FP&A Guy

Alrighty, I'm live. We don't start for a few more minutes, but if you're out there and you can hear me, let me know. So we'll start at the top of the hour, but I wanted to just jump in a little early. So if anyone's out there, they can just let me know they're out there. We got about 5 more minutes before we'll get started though, but if you have any questions, feel free to let me know. And I'm excited to have this session today. So I'm going to drop the link where you can download the files. It's a zip folder. Hey, welcome, Wayne. Anyone who needs it, that's where you can download the file for today. We're going to be— I'm going to be showing an example in Copilot. This is one I've shown before. It was one I shared with Mario Vasquez at, uh, for AFP, and I also used it elsewhere. So if you've seen it there, you'll, uh, you'll get to see something that I've done before. But we'll get started here in a few more minutes. So I'd love to know where you're coming from. If you're out there, just go ahead and let me know. Always love to see. You could share anything you're hoping to learn today, and we'll get started in about 3, 4 minutes here. We're not quite to the top of the hour. So welcome Bradley, Wayne, I know you're out there. Looks like I see others as well. There may be some of you on, uh, LinkedIn, YouTube, or, uh, StreamYard. So we got Bradley from Arkansas, perfect. Now we have Wayne. I can't remember where Wayne's from, it's slipping my mind, but I'm gonna share my screen Just while we wait to get started. And there you're gonna see the website for today where you can download the file we're gonna use. I'm also gonna start us off with just a fun little video that is appropriate for this training. So we got a couple more minutes before we get started. All right, I'm just pulling it up here on the screen. There's the one I want. All right, we'll get to that in a minute. Welcome, Khalid from Egypt, Cassio from Brazil, and if I said your name wrong, I apologize. Jeremy from Sacramento, welcome. Bradley, I'm excited to share a few things. Uh, Jessica from Alaska, love it. So we got a few more minutes before we'll get started. But while we're waiting, I'll tell an Excel joke. Does anyone know what the VLOOKUP said on its online dating profile? You have to be a real Excel nerd to know this one. So you'll— if you get it, you're a real Excel nerd. If not, it's okay. You can just know I'm an Excel nerd. So the question is, what did the VLOOKUP Put on its online dating profile, seeking an approximate match. It's true. So if you know your VLOOKUP formula, if you want an approximate match, you have to type true. So there you go. Didn't say it was a good joke; just said I'd tell it. We got about another minute before we'll get started here. So I'll just give a minute for everybody else to come, but please keep letting me know where you're coming from as you join. Would love to hear. We're going to go through a Copilot example and just share a few of the things I've learned. What I'm going to say before we get started here is Copilot keeps getting better and better. I mean, I know Claude's been really good, but I continue to see improvements from Copilot. So excited for its future. All righty. Sorry, the video and audio is not working. You're the first to say that, David. Everybody else has said they could hear me, so you might want to try logging out and logging back in and see if that fixes it, or try it on LinkedIn. But if anyone else can't hear me, let me know. David says it's not working, but I believe everybody else has told me it's working, so if something's changed, let me know. All right, we're top of the hour, so why don't we go ahead and get started? So we're going to be talking about Copilot today, and I'm going to share an example around scenario modeling with Copilot. But before I do that, I wanted to share just kind of a fun little video that, uh, in the 1990s Microsoft made an ad, the Excel Elevator ad. It was really famous about how their spreadsheet was different. If you haven't seen it, just look up Microsoft Excel Elevator ad. They redid the ad last year for Copilot. And I, since I thought since we're talking about Copilot, I'd have a little fun and start by sharing the video. So you can see a classic Excel ad just got a 2026 update. So give me one minute here. If you can't hear it, let me know. But I do have sound on, so everybody should be able to hear it. And here we go.

Did you see Wilson's email?

Yeah, I'm on it.

What, what is this?

It's Copilot in Microsoft Excel. Now hold still, I'm about to perform a miracle. Stuff that takes hours, we'll do it in minutes.

My spreadsheet doesn't do that.

Okay, I admit it. It's pretty neat. But there still isn't enough time.

Don't panic. I'm just going to have Copilot build another report. My spreadsheet

doesn't do that. Oh, look at this new ask from Wilson. She wants a price comparison. Give it up. Look, look, let's just say that your PC got run over by a bus and that you're lucky to be alive.

Oh, ye of little faith. Pull price information from competitors listed on Sheet 2. With a few clicks, our superhero Fixed errors, incorporated relevant data, and cleaned the whole thing up.

My spreadsheet doesn't do that.

I can't believe it. You did it. This is a miracle.

This is Microsoft Excel.

All right, so a little bit of fun. If anyone's seen the 1990s ad, it was kind of a remake of it where Microsoft was saying, hey, our spreadsheet is different. All right, so let's jump into things. I see several others have let us know where they're coming from. Kaleem, welcome from Richmond. Um, we have Karmath from Saudi Arabia, David from Serbia. So keep letting me know where you come from. Let me know your questions as we go. If you haven't downloaded the file, you can get it on my website. If you go under educational resources, Excel files will be the first one right here. You'll have 3 files that it will download. I've also dropped that in the chat for anyone that needs it. And there'll be a template in there. There'll be a Word document with the prompts I've used for this. And then I believe there's the final file as well. All right, so let's go ahead and get started. This one I'm going to do a little bit of a presentation, so it won't be as hands-on as some of mine, but we'll definitely do some work in Excel. And to follow along where I do do some prompting, you will need the paid version of Copilot to work in Excel. So just be aware if you're hoping to follow along, you could also follow along with Claude or ChatGPT if you have that paid version. There'll be some subtle differences, but much of it will still apply. All right, so we're going to do a little bit of a case. I'm going to set up the case we're using here, then I'm going to show you example 1. We're going to do prompt only. We'll talk a little bit about prompting strategies. We're going to do example 2, which is what I call best practice prompting. Then I'm going to talk briefly about what I call context engineering, or kind of how to give it context beyond prompts. Then we'll do example 3, and we'll have some key takeaways. So that's what we're going to go through. Ask questions you want. What the scenario is, there's a link there if you want to download the Excel file if you haven't got it. You can do that real quick with your phone. I'm going to skip over that as I think I've showed it several times, so everybody should have that. So the example we're going to do is a lease versus a buy. We're a finance audience. I'm not going to spend much time here, but this example, we're going to have 3 different scenarios that we're going to have Copilot model. We've all dealt with some of these where, hey, should we buy this product? Should we lease it? Right? And there's pros and cons. Leases have less capital. They sometimes are off balance sheet. You may not have to deal with all the maintenance. They're quicker to deploy. Buy means you own it, you get all the depreciation for sure, but you have a large cash out front. You're on the hook for all the maintenance. It can impact your balance sheet, disposal costs, all those type of things. So the situation— and this comes from Mario Vasquez, it was a real situation he dealt with— is we have in Iowa, Cedar Rapids area, you can see here on the map, they need to replace their broadcast tower for TV. So what had happened is all the equipment had to be replaced. They changed the, the, uh, frequencies they were using. The government in the US changed it, and that meant they had a couple different options. They had an expensive lease. They could choose to continue that lease and just update the equipment. They could buy a new tower similar to that one there, or they could lease a tower from somebody else. So those were the 3 options they had in this situation. And you can kind of see the coverage area, you know, give you an idea of what we're dealing with. And so I want to list the assumptions. I could have included all these in the exercise, and Copilot can handle them. I chose to keep it relatively simple because I want to teach some principles versus showing how far we can push Copilot. So this course is really about getting some ideas that will improve whatever you're doing. With Copilot, not hey, how exact is this of a real-world example? So there are some shortcuts. So we didn't consider time value of money, but that would have been a pretty easy thing to add. We didn't have AI analyze the tower specs. So if this was a real decision, you would have the financials, but you'd also look at how powerful a tower is and all those things that aren't. Quantitative, they're qualitative. They're not in the you know the financials. The instruction sheet included all the cost estimates. You may not have all of them up front. In this case, we had them. We didn't include inflation assumption for recurring costs. We kept it flat since the recurring costs were pretty much the same across all three scenarios. There was very minimal difference. We figured why make it harder than it needs to be. And as I mentioned, I'm going to use Copilot for this. All right, so what I want to do now is we're going to go into Excel. I'm going to go into Copilot here, and there's 2 things I'm going to do. So if you have Copilot, you can go ahead and open it with me. I have it here on my menu bar. If not, you may have it down here. You can dock it. So if I right-click If I right-click on— I'm sorry, not right-click. If I click on the dropdown, I can move it out of the ribbon, and you'll notice it's now down here. It shows up on the bottom of the screen, and I'm going to remove the comments so it's easier to see. So let me just turn those off, and you'll see it right there, very bottom. I'm going to right-click on it, and I'm going to dock it Because I like having it in my home ribbon. Not dock it, I meant I put it in my home ribbon. So let me right-click again, move to top. So now I have it here. All right, so we're going to go ahead and open Copilot. And the first thing I'm going to do for this very simple example is I am actually going to click on these 3 dots. You have this thing called personalization. I'm gonna turn— I'm gonna delete that for a minute. I'll explain why when we go further along. We're gonna assume we just have a blank sheet of Copilot, and all we've given it is I have given it the prompt, which basically says, hey, it's 2017. Here's the example. There are 3 different scenarios. We have 3 options. We give it option 1, option 2 here. Option 3, we give it all the prices, the recurring costs, headcount is this, station expenses this. I've given it the details in there. You can see it on your screen. We tell it how much the equipment will cost. So what I'm going to do is I'm going to start in the Word document that went out. There's a very basic prompt. All I'm going to give it is say, hey, I have the following instructions in a tech sheet on the sheet labeled instructions. And I need your help building me a model using the instructions to guide you. Can you build a model for each of these 3 scenarios mentioned in the instructions and a summary comparing the 3 and tell me what I should do? So that's all I'm giving it. Pretty simple, simple model. We're going to drop it in here. We're going to have editing on. We're going to let it select what model it wants. So it's an auto. We're not selecting the model ourselves. And I'm going to kick that off while I go back to the PowerPoint and share some examples. You're welcome to kick it off with me as well and see what you get. All right, so I'm going to give it 1 minute because it will probably come back and ask me a question, and we'll just answer that question real fast before we continue. So it's telling me I'm reading the instructions, I'm extracting. We'll see if it has any questions here. The first read encountered is a non-text shape. I'm narrowing the inspection. So it takes it a minute to realize there's words in the box. I may have to specifically tell it. Sometimes it doesn't pick up. That's why I said instructions in a tech sheet is you'll find Copilot sometimes struggles a little bit with tech sheets. So we'll give it a minute and I'll come back in a second. I'm going to go back to PowerPoint here and we'll, we'll check on it in a minute. All right, so let me Get this back up. All right, so we saw the prompt I gave it. Here's what I got the first time I did this. So I want you to think, if you put a summary sheet— this is the summary sheet— like this in front of your manager, what would you probably— the response be? What I will tell you, the numbers were right. So this is the, the right calculation. So did the right math, but probably not the easiest summary to read. Would love any thoughts. If I— if you put that in front of your manager, what would they say? I tell you what most my managers would have said: go back and clean it up. So that was prompt— that was try 1. Here was try 2 with the exact same prompt. Now look how different that is. Prompt 1, different header. Uh, they both have a story in brief, but different wording for that. The financial comparison is a little bit different. One included a coverage variance, one did not. One included a graph, which— does the graph tell you much when they're all kind of the same? Only a little bit. This one did a clear recommendation. This one didn't give any words to its recommendation. So what you're seeing, I ran that exact same prompt twice, is very different results. But what's important to note, okay, so it's building right now. So we can see it. I gave it that prompt. It didn't ask me any questions and it's building. So you can see I have a summary sheet right now. It's not done, but notice it looks different than the other one. This one doesn't have a table. It showed the annual cost, so it took a different approach than the other times, but same numbers: 8.9, 7.5, 7.8. So what what we're finding now when we give it a modeling case, if you can give it all the assumptions, the math is usually pretty good. The math is often right. You do have to check for hard-coded values. You do have to know what you're looking at almost every time. If I go through this in detail, I'll find somewhere where it hardcoded something. Not always, but almost always. So that's something you'll have to check for, is say, did they do any hardcodes? I mean, divided by 12, that's a pretty common one. I wouldn't be too worried about that. Otherwise, you know, looks pretty clean on that page, but it's still building. But again, looks quite a bit different. So what we're finding, if you give Copilot a basic prompt, you'll get something decent and it may even do all the math right. It's not going to be presentation ready. It's not going to be in the format you want. You need to check it. Right? So there we go. It just finished. I can hit done. That would be build number 3. You notice it did everything in the months in columns or in rows. I might have wanted them in columns. So, but that's what happens when you give it a very basic prompt. So if we start there, Copilot is okay, not— I wouldn't call it good. So we started with that. The next thing I did, we're going to go back to the presentation here, the slideshow. So here's a little bit of the basic prompt. As I mentioned, each attempt has widely different layouts. It will make— it made some different assumptions on key elements. Formatting, look and feel is often very different. Dates were in rows, categories in columns, not necessarily what I want. Final numbers tied in all 3 examples we just looked at. They had the same number. So that's what we got with the basic prompt. Now, many people say, hey, it's all about prompting. If I give it a good prompt, I'll get something good. The reality is today, if you know what you want, you can ask AI to help you build the prompt. I know several people. I built one for a training we did for a prompt using the bridge method where Claude helps me with— I tell it what I want, it helps me write the prompt. Copilot's created an agent. I believe they have a tool that will help you write prompts. So it's getting to the point now where AI can probably write a better prompt than us. These are some of the common prompting methods. I'm not going to spend much time here. I want to get into the examples a little bit more so you can see the journey. But, you know, structured prompting— you give it the role, the context, the task, the format, the constraint, the examples. That's probably my favorite in finance. Not a big fan of few-shot prompting, which is, hey, you give it several examples, you give it some prior examples and just let it work. We usually need to be too specific in finance to do something like that. Chain of thought is a good one. Always ask it to provide its reasoning. I like to kind of combine structured chain of thought and depending on the exercise, prompt chaining where I break it down. The more complex, the more you want to break it down. So what I did is I had AI help me write a best practice prompt. I told it who I was, told it where the instructions were. I gave it constraints. I gave it all the information that were in the instructions. I gave it exclusions. I said, here's the task. I want you to build an annual cost model. I said, prior to building, please validate all assumptions and build a plan for me to review and approve. So that's the second example. Much better prompt than the first, and so let's see how it did. But before we did that, let's see if I can tell another Excel joke. Does anyone know why the spreadsheet was constipated? This one gets me in trouble. Anyone know why the spreadsheet was constipated? Because it could not budget. Didn't say it was good humor, just said it was humor. All right, let's keep going. So here's what I got with the best practice prompt. When I ask people— and I wish we could come off mute, we're not in that environment where we can with the LinkedIn Learning— but put it in the chat, is this really that much better than the basic prompt? This cover sheet. If you look at the details, what I found is even with the best practice prompt, it still varied quite a bit. And we could even give it this prompt and let it build if we want. So what I could do, we'll go ahead and do that so you can kind of see, is we're going to come back here. Ah, Gary, full of sheet. Close. It was constipated because it couldn't budget. But I like full of sheet. That's a good one, Gary. All right, so we're gonna go back to our template here, and I'm gonna delete the basic example we had, and we're gonna give it the best practice prompt. So I'm gonna delete— wait, no, I don't want to delete the instruction sheet. I want everything but the instruction sheet. Not on hide, on select, on group. Let me regroup that. We'll delete these and you guys can do it with me, but you can see this one looks quite different than the other one we looked at. So every time we get something quite different. That's one of the big complaints people have is how do you make AI consistent? That's what we're really going to show with example 3 here in a minute. I'm going to clear this and I'm going to copy from the Word document. My best practice prompt, which is this one right here, all the way to here. So we'll just, uh, put it in here. Oh wait, I, uh, I think I added one extra line. I didn't mean to include the info included. But let's see if it gets it right. We should be okay. Should be able to figure that out. I should have just— oh, I included— yeah, let me— I accidentally included some of the best practice. Let me start that over. It should have been to just right here. I included a little more than I wanted. I included some of the context that we'll do next time. So let me kick that off. There we go. All right, so we'll let that work, and I'm going to show a little bit of the deck. We'll come back and check on that in a minute. All right, so let me just bring that back up. So that's what I got with example 1. Not, not much different than the basic prompt. Same answer: 7.8 million, 8.9 million, 7.7 million. Exact same math, very similar layout. If you look at the other sheets, it may be a little better with the best practice prompt, but not materially. Try number 2. This might be the worst of all of them, in my opinion. I mean, the top part where the recommendation is, I don't know why they're using like 4 different font sizes. You'd think it would be consistent. You know, leadership view of undiscounted nominal costs over engineering and coverage are excluded. Awkward reading, but okay, fine. You know, and then key findings. It doesn't even give a recommendation, which we asked for a recommendation. So I would call that one again, okay. So what we're finding is even with really good prompts, you're probably not getting consistent or much better results than a basic prompt for many of the things you're doing in Excel. Whether it's Copilot or Claude. So let's now— we're going to do a 3rd example. I'm going to go back into Excel for a second and see how the 2nd one's coming, see what questions they have for me. But we're going to walk through several things we're going to do to improve what we get. All right, so it included the COVID sheets as requested. It had included years for the model. You know, the tries had a lot of variation. So like I said, slightly better. That's a little bit about that. Now we're gonna get into what I'll call context and what, what that really means when you're working with AI. All right, before we do that, let's go back here and ask the question: which option 2— which should option 2's 4,000-month lease begin? It should be October 2018. That's when we want this to start. So that's the only question it had. I know in one of these I might have had a wrong date somewhere, and it probably picked that up, which is good. So it's validated, saying, hey, here's all the assumptions. I reconciled your brief. No workbook changes have been made. Then it tells me, here are— here's how I plan on building it. Here's your assumptions. Here's your monthly model, your annual summary, the core escalation logic. So how it builds the escalator, which makes sense to starting rent times 1 plus the escalator raised to whatever year it is. And so I'm going to say go ahead and build, or just approve, and we'll let that build. So that's going to do a best practice one. Well, it's doing that. I'm going to come back to our deck and really talk about how we get something repeatable. So when we first— AI came out 4 years ago, all we heard is the value of the prompt, right? The whole idea of prompt engineering. Then we kind of did a journey. I'd say a year ago, 6 months ago, as Claude got a lot better, we started hearing a lot about context engineering, which— what context engineering is providing details along with the prompt, sample files. instructions, skills, personalization, and workbooks. A lot of things like that. Now we're hearing harness engineering, which is some of how do you make it deterministic? Maybe you're using scripting. It's all the guardrails you're giving these agents so they can do a better job of completing multi-step tasks reliably. That's been the journey. For this example with Excel, we're really focusing on context engineering. So as we go beyond prompting, there's several things we can do in Copilot and in other places. Instructions, personalization. If you haven't personalized your AI, you should. So you're aware, and I'm going to show personalization now, and then we'll come back to the others on the list. In all of your tools, in fact, if you go to the 3 dots and and you click on Personalization, you can do it with Claude, you can do it with Copilot. You can give Excel specific instructions it should follow every single time you use it. I'm going to show you mine in a minute. You can do the same with PowerPoint, you can do the same with Word, you can do it with Outlook, in addition to doing it with ChatGPT in general. So these are more specific instructions for the application that you always want it to follow. Give me a minute. My Excel has— there we go. It froze for a second. We're back. It's working on this example. We'll see how it comes out with this best practice prompt. But instructions, projects. Projects are just a place where you can save everything so you can save all that context, create memories. For different things you're working on. Skills— skills are nothing more than instructions written in a language that is easy for AI to understand. They use a markdown language, which makes it easier for the LLM to understand. Copilot has them. I'll show you that in a minute. In fact, I'll do that now. You can manage plugins and skills right here. You can create your own custom skills. That may— that's not in beta, that's general release now. You have finance skills you can turn on, you have visualization skills you can turn on. If I click here, there's chart— there's a chart design skill now. If I check under finance, there's a number of different skills that are default that you can use to assist you in your work. They use some of FMI's Financial Modeling Institute's examples. For those skills. All right, so that's an example of what I mean when I say skills. There are a couple other things I'm going to show in here, but before I do, let's see, it finished our model. Let's just take a look. So here's the assumptions, pretty clean. I'd have to look through it all to say, hey, are there some issues? You got to check every little thing to make sure, but it looks pretty decent. October 1st, September 30th, that's right, that's right. For the current, the office expenses, the headcount looks right. 2 point— I can kind of spot check this as I know most the numbers and everything on real Looks right. We'll look at the amounts came out when we look at it. Okay, so here's one thing that's different. This time it did the monthly model where it did all three examples on one page. I don't know about you, but I don't want to have to go over to AN and try to figure it all out on one big page really long like this. Probably not the way I not the way I like to model. So again, it's because we didn't give it enough context. Gave us an annual summary, and it did a gap to next best by year. Yeah, I don't know how helpful that is. Lowest cost, and what it did is I think it did lowest cost for each year, so it's giving you a different model each year. Doesn't even tell you the model. Again, not real helpful. And then I have the executive summary, a decent This is easy to see, 10-year total cost, but it's not right. All 3 numbers are showing $0.0 million even though they have the right amounts. So that's an issue. You know, this here, don't know how helpful it is. So again, best practice prompt, not great. All right, so that's example 2. So now we're going to go to example 3 here. So like I said, connectors, MCP, spreadsheet agents, these are all things that help you go beyond prompting. In example 3, here's the things I did. We're going to do all those and then I'm going to show you what I got. I included personalization in Excel. I'll talk about the type of things you should include there. I included workbook rules to follow. I included a sample file. I used the plan mode. And not every time, but often I use Claude Opus 5 because I do think Claude in general does a little better job with models. Depends on who you ask and what you're building, but in general it does a better job inside Copilot. I think Claude as a whole, if you're using the Claude tool, does a little better job on modeling. All right, so let's walk through what each of those look like. Okay, so if we come here, These are my Copilot instructions starting right here. So I'm going to put them on the next page. And I'm just going to— so right here they're highlighted. I'm going to make this a little bit bigger. So you're just seeing that. And so first thing I tell Copilot as an example, all tables you create should contain short descriptive name. And be prefixed with table. For example, table revenue or table expenses. You should use tables to store raw data when possible. There are situations that may not make sense. Never merge cells, always use center across selection. I don't know how many have seen this. I think it's in this version. I'm going to open up Excel again here. I think it got closed, so I don't know if it finished building, but they've now added Oh, let me open this up. Yeah, okay, we looked at that final one, but they've now added Center Across Selection is now on your home screen. So that's a big deal. That's something that took them forever to, to fix. How many of us knew if you wanted to center something across selection, you had to go here and you had to go down to Alignment and you had to manually find it. Now it's on your home screen. All right, sorry, I digress. So I say always turn off worksheet grid lines, start new worksheet layouts in B2 by default. When building models, use blue font for input cells, black font for output. And I give it do not hard code. So those are all the instructions I give it. That's the first thing I do. So I'm going to take that. We're going to do the final example here. So I'm going to go back to Excel. I'm going to delete this, and we're going to build for the third time. Actually, I'm going to use a different file for this third one. I am going to make— let's see, I think I need to make a copy of this one. So I'm going to open this. Give me one minute here for it to open. Here we go. All right, so, all right, gotta love when Excel doesn't want to open. So we're gonna try this again. Let me close Excel. Appreciate the patience. If anyone has a good joke, let me know. All right, so I'm going to open this one here. Uh, looks like— okay, yeah, here we go. It's now open. Let me close this. Ah, it's got 2 versions open, so I'm just gonna kill Excel altogether so we don't end up with a problem and reopen it. You guys get to see what really happens with Excel, right? We all— we've all dealt with it. All right, there we go. Let me open it again. So the first thing I did is personalization. If you're not sure what to put in your personalization, well, I don't know, Gary, what's orange and sounds like a parrot? Let me know in the, in the chat. I am curious to hear this example. All right, so we're gonna delete— this was example 3. We're gonna recreate it here. So let's come here. So as I mentioned, the first thing we did is we finished putting in our personalization, which we should be able to see. I didn't save it, so let me drop— let me copy it from here. If you're wanting to create your own personalization, tell Excel how you work. You could even ask Copilot a carrot. Well done. You can even ask Copilot to help you write. Say, hey, I want to create, customize my Copilot, create some instructions. Can you ask me how I work in Excel and write some for me? It's better than nothing. It gets you started. And everybody will have something different, but there's mine. So that's the first thing I'm going to do. So I, I gave it some context here. Second thing I'm going to do is I'm going to go to Auto, I'm going to go to Claude, and I'm going to use Opus 5. So that's the second change I've made. Third, I'm going to change editing to plan mode so it has to build the plan right here. And then if you click here Notice you can add your own custom workbook rules. What I'm going to do is I'm going to go to a blank sheet for 1 minute and we're going to add some workbook rules so you can see what that looks like. So come here, click here, create workbook rules. So here's an example. This is the default. You'll want to remove all these e.g.'s. So that it knows. Yes, Bradley, great question. Bradley asked, so I'm going to step back for a second. Do you have to set up the personalization? And let me just show his question so everybody can see it. Once, does Copilot keep them? Yes, it will keep it for as— unless something resets, like your account gets reset or something, it will always keep it. I didn't have mine in because I wanted to show the basic prompt without them. Otherwise, I would have got different results. So what I want to show is as you layer on more and more of these things, you get better and better results. So you've done your instructions, you've chose the model you think will do the best job, and that's a little bit of guessing, that's a little bit of experimentation. Everybody has different opinions. You couldn't just select auto if you need to. Third thing is workbook rules. So here it says replace these examples with rules that tell Copilot how to format data and sheets in this workbook. So in this example, format currency as this, use bold headers, prefer XLOOKUP over VLOOKUP. Our fiscal year starts in July, right? You give it all that information. Name new sheets with the format of whatever. All right, so that's the blank one. What I did in my example, if we look at rules, and I overlap some of these with my Copilot instructions. And I'm okay with that because it doesn't always follow everything we give it. So I gave it to it twice. I'm going to make this a little bigger for everybody. And so I said, hey, I want the dates to be in columns, categories and account lines should be in rows, the header section should be the first 4 rows, and all 4 rows should have a dark black background or blue background. And the next one that I told it should use white font, I kind of repeated the other one. I probably could have combined those, but I continued to give it several different things I wanted to do. For example, use a white background for all cells unless explicitly stated otherwise. So those are the rules I gave it. So that's the third thing— fourth thing I've done. So if we go back here, I'm just going to go through them real quick. The 4 things I did: I did my personalization, I selected the model I wanted, so I said Claude Opus, I put it in plan mode, I added workbook rules. Gary, I would say if there's conflicts, it will usually follow Copilot first, but the reality is it doesn't always follow everything you give it. It is AI, it is probability, so it could vary some. But I believe my understanding is personalization has priority in the way it's structured, but that's not a guarantee it's always going to follow it. Then the last thing I did— so we've gone through 4 things before I give it the prompt— is I gave it an example. So I said, here's an example of what I want. I gave it a summary sheet, an executive summary. I deliberately only gave it 2 scenarios instead of 3. But I showed it, hey, here's how I like my headers to look, except for I didn't start in B2. So you'll see, does it follow the example or does it follow the Copilot instructions? I kind of showed it an example of what the, the assumptions should look like. I gave it my scenario data. So now I've given it basically 5 things: personalization, selected my model, I've told it to use plan mode, I gave it workbook rules. I gave it an example of what I want a scenario to look like. All right, so now I'm going to go ahead and give it my prompt. It's basically the best practice prompt with just a little bit of detail to explain that there is an example sheet it should use. The prompt here tells it, hey, I've given you an example. That you should use. Otherwise, it's basically the same as the best practice prompt. So now we've brought everything together and all the context I've mentioned, and we're going to kick that off and let that work. Before I— we'll go ahead and kick it off now, and I'll probably pull up the example, but I'm going to go back to the deck for a few minutes. Okay, so actually, no, let me— well, that's working. What I want to do is I'm going to show you what I got last time I did this. Well, it's building. So this is one with everything I've mentioned. So first thing it did is it built a separate sheet where it did the monthly detail. For each of the models. Okay, great question, Leah. So the sample file right here, it's the sheets named Example. In Copilot, you can't upload an Excel file. You could upload a PDF as an example, but you can't upload an Excel file within the Excel agent. So I put my example In the workbook. If I was doing this in Codex or Claude Code, I could have uploaded the Excel file and said, here's my example. I've done it that way as well. But here I gave it the example within the workbook I asked it to work in. So hopefully that answers the question. If not, let me know. But Leah's question, you can see it right here, everybody. So I want to give it the example before I give it to the prompt. Otherwise it will say, hey, where's your example? So I put it in the workbook and I told it in that prompt, hey, see the workbook. All right, so let's look at what we got. So if we look at the example, I got a cover sheet a little different than the example I gave it, but honestly probably better in that it gave me a sheet index. It said, hey, here's what's on each sheet, and it gave me a purpose. But you can see it started in B2, which I asked it to do. It did 4 rows in blue. It did a white background. It didn't do background anywhere else. So for the most part, followed the example, followed the instructions quite well. Executive summary, very close to what I would have expected. Green linking to other sheets. Again, started in B2, showed the example, you know, gave the total, did a check to make sure everything was tying. So built in a check. And then it gave the recommendation. So it's cleaner, it's simpler. I could have asked it to do graphs, but it, you know, it basically followed my example, but it also looked at the instructions where I said give a recommendation and it added a recommendation. So good job there, much closer than what we were getting before. Assumptions— it didn't use the orange background, but it did do a note like I asked. It did do the unit, the input. It kept them. You could see it followed the similar formatting where, hey, I have my model assumptions, then my general, then my tower and lease, then my recurring, then my one-time. Very clean, easy to update. Is it perfect? No. But without, without skills, without deterministic, without lots of examples— and even then it may not be perfect— you're not going to get everything you want. But much, much better. Now let's look at Scenario 1 again. Clean, has the expenses, the one-time expenses, total expenses. It did it by year like I asked. It did formulas. What I did find when I built this is in one or two of these places it missed a dollar sign. There was a hard-coded number. It wasn't perfect. One of the times the math was slightly off. Again, you could write a skill to review things. You can do agents. There's things you can do. But what's nice about doing AI in Excel versus, you know, code or whatever is you can always see each of the formulas. Now you can do it in Claude Code, then open the file and audit the formulas. But having it built in Excel, whether it's Claude Code or an Excel agent, is a nice way for transparency, which we all like in finance. All right, so if I go to the next example, again you can see almost the exact same layout except for it had some different expenses which it combined all together on the executive summary, but it only included the expenses for each page, which is fine. I probably would have gave it, um, instructions if I did it again and specifically said Hey, I want the exact same expenses on each sheet. I showed the same ones on my sample sheet. I'm not sure without instructions it really picked up that level of detail. And then it did the monthly detail on the back where it had everything, all the options I can go through and kind of look at all the formulas and say, hey, am I good with it? Does it make sense? Do I have issues? And so there's an example of going and see, you could see it again. This one here, we'll let it build. Option 1, renew. How should the renewal rent be built after the current lease ends? Continue same lease. Option 2 and 3, pay the current site lease and the new site in parallel. Confirm the overlap. Yes, that's intended. Okay. How should the annual periods be deferred? We'll go with fiscal years. One-time costs are stated as incurred before— where should they land? I'm hoping everybody can hear me. If you can't, let me know. I lost my mic for one second, so I apologize. So let me know if you're able to hear me. It should be back on now. So just— okay, perfect. You can hear me now. Thanks, Amanda. My internet just went out for a second as well, so I'm having good luck. Perfect. I'm glad the audio is back. Sorry about that. All right, so let me just recap the last slide one more time since I lost things for a minute. I was just saying how I set up Copilot, the personalization, the work. And I have no idea why my browser dropped my mic again for a second, but I'm gonna keep going. Hopefully it doesn't do this anymore. All right, so yeah, it should be back now, Jeff. Sorry about that, it went out for a couple seconds. Hopefully everybody can hear me again. I apologize, I have no idea what's going on with the connection. So, all right, so here's my takeaways. Every model had issues. None of them were perfect. The best results almost always come with context. Context. The Excel agent, even with an average prompt, can probably produce the right calculations, the right results. Sometimes the formulas may be overly complex. It may not always be in a format you want. The scary thing is it's almost always polished, but not always right. So you need to be really careful for that. And the reality is prompt— what we used to call prompt engineering, it's given way to context and harness engineering. You want good results, you need to go beyond a prompt. It's just not enough. All right, so let's talk about what actually happened in the case. In the real case, they purchased the tower, which was slightly more expensive than the least expensive option, but it had better household reach. Which that gets back to, right? As good as AI is, we as humans need to be in the loop and we make that final decision. There's still human judgment involved. And it helped their negotiations with the big tower owner, gave them a huge savings after year 10. So what it did is having another option, they were able to go back to that tower owner on other towers and renegotiate. So that's what actually happened. Funny thing in the case as well, just you know, when they bought the tower, or the tower they got, it came with farmland. And so every year they get a check from the farmer on the land where their broadcast tower sits. In the, in the real case, that's the real case that they actually get. So last thing I'll share, I do a lot of Excel training. I have, uh, my FP&A Ultimate Course Bundle. If any are interested, you use the code webinar, you can get 30% off on that. Um, there's also— oops, that's in there twice, so let me— this shouldn't say course bundle, this is my Excel one. Nope, that's the wrong one. Just my course bundle here. If any of you want to follow me, there's my website, my LinkedIn, my newsletter. What I'm going to do is I'm going to go ahead and tell one more Excel joke. And then we have a few minutes left. I'm gonna open it up to questions about Copilot, about Excel. You guys can ask any question you want. We got about 10 minutes left. So the joke I'm gonna tell— what did the statistician look for in a spouse? So the question is, what did a statistician look for in a spouse? A normal distribution. All right, so I'm going to let it proceed while we chat. So feel free, any questions, I can bring back up that Excel in a minute, but I'm just going to open it up to any questions anybody has. Amanda, I'm glad you enjoyed it. I hope everybody found this beneficial. If you didn't, I'm sorry, but I'm not sorry. Not much I can do about that. But I'm opening it up. Any questions anyone has? Copilot Excel, even if somebody has an FP&A question, feel free. I'll hang around for a few minutes. What would you say are best practices for auditing an AI-generated financial model? Um, great question, Jeff. So first, a few best practices. One, to help prevent reviewing every calc, every tab, are there places you can put checks in? So can you have it summarized? Like if you're giving it a data file with numbers, can you have it have a place where it checks against that so you can check that? So that's one thing to do. Always build in checks. 2, always have AI do a review. You can build a skill for that. It could be agents. You can do it via a prompt. But you probably want it to review itself. I say at least 3 times because it will find mistakes. From there, you still need to review everything. The more important your analysis, The more thorough of a review you need to do. The more complex of a model it built, the more thorough the review. The lower the risk, the more simple the model, the more you're just worried about the output, you may not have to check every calc. That's a judgment, but I would until I'm quite comfortable. It's a general rule: use an agent to review itself, or use a prompt or something. Have AI review itself. And then review it as well. Build in as many checks and calcs as possible. Um, Dale, I I chose to use Copilot just because it was it was easy to explain. It has workbook rules. Um, a lot of times I train people with Copilot, but you could have just done it with Claude, and I could do it with Claude. There's But those are the reasons. The only thing that's going to be really be different between the two outside of Claude generally gives a little better results. But as far as layout, they're both going to have personalization. They both can take an example. Copilot has a plan mode. Claude, you got to specifically tell it. Copilot has workbook rules built in. Claude does not. Claude skills are a little better implemented than Copilot. the way you can add reference files. What we're finding is if you— we did a study. If you're a small shop, you have very high Claude usage, and the bigger your company, the more and more your Claude usage comes down. If you're a Copilot shop, if you're a small company, you have very low Copilot usage. Bigger your company is, the more likely you are to use Copilot, which speaks to the fact that most large enterprises are Microsoft shops. Hopefully that, that helps a little bit. Khalid, you're welcome. I'm glad you enjoyed it. Bradley, glad you enjoyed it. Gary, thank you. Wayne, thanks. I'll hang out for another— I'll give it about 2 more minutes, and if there's no questions, we'll go ahead and just end a couple minutes early. All right, let's give it another— leave it till, uh, 5 after or 5 before the hour, and then we'll, uh, shut this down. So we got one more minute if anyone has questions. And while I'm waiting to see if there's any more questions, let's see if I can do one more Excel joke. What does every newborn spreadsheet need? So the question is, what does every newborn spreadsheet need? Formula, of course. Thank you, Jeff. I'm glad you found the session helpful. And if any of you, if you need, uh, training in your companies or you need help with FP&A, please reach out. Check out my website. I love doing these type of things, but I also like learning it or, uh, earning a living. So welcome, Dustin. We're just wrapping up, so you'll be able to watch the video if you want. Literally was just probably 30 seconds away from Shutting things down, Dustin, but it will be on YouTube. It will be on LinkedIn. If you registered on StreamYard, you can go back and watch it. So thank you everybody for joining. Really appreciate it. Everybody have a great evening. Go do great FP&A work or great day morning. I think for you, Dustin, it's probably nine in the morning if I had to guess. No, actually probably seven. Early for you. But thank you everyone, appreciate it. Have a good day.