This Excel AI Agent Built Me a $1M Financial Dashboard in 10 Minutes (LIVE Demo)

0:00I spent years inside Microsoft Excel, building models, forecasting revenue, cleaning messy data. I hated it. Every time I thought there has to be a better way. Then last week I came across Shortcut. It's this AI app that plugs into Microsoft Excel, Google Sheets, and suddenly the tedious stuff becomes magic. So I invited the co-founder Nico onto the podcast. we built financial models in real time we explored how shortcut works under the hoods and we talked about how to get the most out of this product if spreadsheets are a part of your daily life this episode will change how you work let's dive in so i reached out to nico and i i I think the DM said something like, this is one of the most impressive AI demos I had ever seen.

1:01Can I use the product? And you said you'd come on to the podcast and share how people are using your startup shortcut. It's kind of like Microsoft Excel if it was built in the future. And I've never really been a huge Excel guy because to be honest, it's been overwhelming to me. the macros. Like, I don't even know what a macro is. Right. Right. So that's why I reached out to you. I just feel like there's an opportunity to, like, I know how valuable it is, but I, I, I just, I know how hard it is to use. Anyways, Nico, welcome to the pod. I want you to show me how, how I could use shortcut to, to make more money and be more productive.

1:54And, and I was hoping you can do so. Yeah, thanks for having me. Super excited. I'm going to just share my screen and kind of get into it. I think I've been doing live demos for some time now, and it's always just the most possible fun way to do this. So kind of like you said, it's like Excel, if it was built in the future. Very intentionally, it's like exactly Excel. You can do whatever you would possibly want to do in Excel, and you would do it here. And we had to recreate a lot of it to allow this. The big difference is that it's also a superhuman agent that can do most of your work. So the good way to think about it is not as a co-pilot for one or two steps.

2:30It's like it will do 90 % of your entire job. And then you get to do other things as that happens. So I can give you some examples. But again, it's just like Excel. And it's not just for creating things from scratch. You can just open up existing Excel files in here and directly manipulate them. So for example, here's a DCF file on Microsoft, which is a pretty nasty model to have to make. And then from here, you can do whatever you want. You can just ask it to be updated. The best way to really use it is to send it off and probably just come back in about 10 minutes. And one thing I want to show you, I guess this would be a technical demonstration of how hard it is.

3:10But I will say, here is this huge DCF, which can take up someone half a day or a whole day to build. Hey, take this. please update it and use the exact template that I want to use Google now. Pull the 10Ks from 2022 through 2024. And do forward projections through 2029. And it's going to do it. it's kind of crazy to watch I think it's fun to watch the first couple times but again you're going to want to just come back to when it's done but Greg the other thing I kind of want to address is like your question which is like how can people use this to make money and be productive and I think the best way to do that or show you that is like to do exactly what I'm already doing right now or like I have to do today so for example this and you'll see that I can come back to and I have multiple shortcuts running at a time this is like dummy data I try to make it look just like our data without giving away sensitive information from actual revenues expenses and its sources and its types and what I need to do for work is build a P &L pretty classic income statement and specifically I need to project it out two years as well while looking one year back and it's going to be for our data room I know you're not a big Excel guy but this is one of the most common financial models you'll have to make on Excel from Rookie.

4:44I will say I'm not a big Excel guy as a contributor to the Excel, but when people give me Excels, I'm loving it. And I want to be able to manipulate it, but I'm scared I'm going to break something. Yeah, that's fair. There's like two actual major use cases right now, or types of users, archetypes. The ones that it's most sticky for are the people who are kind of Excel experts, but it takes hours of work and makes it like 10 minutes. But there's another class of people that I'm learning more and more about, which is they're not super strong at Excel, but this thing makes them almost like Excel gurus pretty quickly.

5:19So we can even do an example together based on what you want to do and see how far you can take it and see how much better it makes you. So here's what that example was for me, what I needed. And right before I do that, I actually will show you one thing. Sam Altman, the co-founder of OpenAI, just said that it is the era of the idea guy and he is not wrong. I think that right now is an incredible time to be building a startup. And if you listen to this podcast, chances are you think so too. Now, I think that you can look at trends to basically figure out what are the startup ideas you should be building.

5:54So that's exactly why I built idea browser dot com. Every single day, you're going to get a free startup idea in your inbox. And it's all backed by high quality data trends. How we do it, people always ask. We use AI agents to go and search what are people looking for and what are they screaming for in terms of products that you should be building. And then we hand it on a silver platter for you to go check out. We do have a few paid plans that take it to the next level, give you more ideas, give you more AI agents and more almost like a chat GPT for ideas with it. But you can start for free, ideabrowser.com.

6:36And if you're listening to this, I highly recommend it. This is kind of an interesting user paradigm. One thing we've learned is people specifically working on Excel aren't extremely good at prompting. And they don't even expect that the AI really greatly understands its subject matter. But when you show clarifying questions, it can make them better and it's like a first magic moment for people who use Excel a lot. So what are the growth assumptions I should use going forward? Conservative, moderate, aggressive. Let's just say I want all scenarios. Let's make it hard. That's really cool. I've never seen UI like this.

7:15Actually, I've seen from time to time a clode or chat GPT will be like, can you refine it? And I love when it does that. But I haven't seen it built, productized like this. It's become one of the magic moments, which I totally did not expect. But users are really not that great at prompting. and I think GPT does something similar if you do deep research or if any of you guys did use deep research a good amount but they almost they like necessitate that clarification but for us we actually want to make it context aware so that the clarification is even better and users have like become much better prompting because of this so I'll do all scenarios let's keep the same exact structure of the template and update the data with the same metrics and charts again I think this is like the most similar to real finance work, for example.

8:03It's like you have your templates and you just want to update them. You don't want to create things from scratch. It's definitely like this is maybe even like the hardest kind of work. So it'll take that on. I'm going to go back to this example. I'll say, hey, I need to build a P &L for this last year of data and do a two-year projected out. this is for my data room for vcs and bankers make it very professional it's always funny to see like how it interprets that um see how it goes and what we can do also is like show you a third version of this and like something you want to try or that you think like would be valuable to you or valuable to your audience and we can give that a shot as well cool i'm also like as you're going through this, I'm just wondering, what is the best way to prompt shortcut?

8:59Is it long prompts, short prompts? You've seen thousands or more of the prompts. What do you recommend to people? Yeah, it's a really good question. In general, I've always liked relatively vague prompts because it forces the frontier models to really get creative and then suss out clarifying questions from you. So specifically because of our clarifying questions, I'd like less for both prompts. I'd like less specific prompts. And then that kind of encourages a little bit more creativity out of the model and then out of your clarifying questions. Cool. So we'll, you know, we'll fill this out as well.

9:33I'll say do this over many different sheets. Cool. And it will get going from there. And meanwhile, I'll check in on, on Microsoft. So what you'll see here is it actually looked for the Google's data, and the 10Ks exist in an SEC database that's super hard to find and extract. But it found these 10Ks and extracted them. 10Ks are like 100 pages of PDF material for public companies. It found all of these Google 10Ks and is starting to extract the data. And you'll see it actually provided a task list here. So its current plan is to read and analyze all the current models, search for and download the 10K filings and extract these, and then start to update the historical data.

10:16and the drivers and the assumptions as well, knowing that the P &L and the dashboard are more formula-driven and will be updated automatically if you can change the source material. That's crazy, man. That's crazy. Yeah, it's pretty crazy to see. We can talk a little bit into the technical details as far as you think that's interesting. But these 10Ks are so big and confusing and horrible that you'll see that we're running into context limits here. You see in the file on the right side. And it's actually agentically deciding, well, let me look at these one part at a time or one chunk at a time.

10:56So there's really no upper limit in terms of how we can allow agents to go over material. I think historically a lot of what we've been doing is rag in this industry. But as agents can learn to start to search for chunks of information selectively and then compact their context as necessary, that is changing dramatically. Yeah, I guess what's going through my mind right now is, you know, Excel has been around for how many years? 30 years. It's almost its 40th birthday. 40th birthday. Like, Excel is a middle-aged person. Yeah. You know, probably a middle-aged man, gray hair, you know, khaki pants.

11:43Yep. And, you know, when I'm looking at this is, I'm like, okay, what does this unlock from a use case perspective that Excel hasn't been able to do? And what are some unfair advantages that people who get onto Shortcut early are going to be able to unlock? That's what's going through my brain right now. Yeah, yeah. So let me tell you a little bit about Microsoft and what we know about Excel. and then what these unfair advantages are. So Microsoft, Excel specifically, I would argue, I grew up using Excel. I started my career in finance, which is not a coincidence for why we ended up building this.

12:23In a lot of ways, I would say it's the best design software maybe ever. It's 40 years of staying power, 2 billion users. The business world runs on it. But what that has accumulated is a lot of things that you have to satisfy for a lot of different enterprises and for a lot of people who are still built 20 years ago for their tech stacks. So even Copilot, which is trying to do what this is doing right now, is forced into helping people use Excel better. It's not forced into like, how should Excel really work? Right? So we really had the chance to just go from the ground up. And instead of doing a single step Copilot for the things that you only want to do in Excel, we have an end-to-end agent that can just do all of your work.

13:09And the kind of fun part about it is that you can import and export Excel directly, so no one would ever even know it's in shortcut. So in terms of what people are using it for, they have their standard things they do at work, and honestly, there's a thousand companies already using this. And I actually think their bosses don't know. I think they have their four or five hours of things they do that have turned into ten minutes, and ideally good employees are doing more of them, but maybe a lot of them are just getting them done and enjoying their free time. And I can tell you what those specific things are but that's I think what the pattern becomes.

13:43So actually by the way here, one kind of really key thing is even if AI becomes perfect, which it's not and I'm not sure it ever will be, it's like an indefinite hill climb, you will have to really, really trust its outputs in order to move forward. As in like if you don't know where things come from, they're almost useless. You have to review its work. So here you'll see that it found these 10Ks from Google, and it can actually cite every single part of the information down to the exact figure, the exact page of the PDF it came from, the exact year the data it came from. So that when things come into your Excel model, you can trace every single bit of it.

14:19It's almost like if you use cursor, CodeGen was very cool, but until you were able to see the diff, you couldn't quite trust it because some arbitrary line in your code would break. So one reason we're not building this directly in Excel, or not mainly focused on that, is that you can't really manipulate the UI. Right now you're seeing that this agent, it sees an error in the calculations, it found it, and it's directly editing it. So we're changing the spreadsheet and the entire front end in real time. I'm happy you said that because it's probably going through everyone's mind, which is like, how can I trust this data?

14:57Yeah, you know, the stakes are a lot lower with like, hey, write this blog post than hey, build a financial model. Yeah, it's just that the finance world until probably right now wasn't ready for this, but this was the same things that we had to ask ourselves in software engineering. It was like, oh, well, I'll never use LLM code. Remember when everyone was complaining about hallucinations in December 22? And then we found out that if you can really observe it, if you can apply the diff, of course humans are still responsible. And if my code doesn't build, I'm the one who gets in trouble. It's not cloud code or cursor.

15:30So what will inevitably happen, whether it's us or someone else who cracks it, is in finance, accounting, FP &A, real estate, wherever, where there's these giant people who use Excel, they'll progress to this role of supervisor, where they're 10 times faster. And if they're wrong, or if there is a bad number in there, of course it's them that's on the line. But they would take that trade-off because it's easier to supervise work much faster than it is to do the groundwork yourself. I mean, ultimately you're supervising work regardless, right? Like either you're supervising human work or you're supervising, you know, agentic work.

16:05And the bottom line is you, you know, I guess the question, it goes back to like, you know, are self-driving cars more safe than human cars? Yeah, it's an interesting thing to bring up. the reason that self-driving cars well I mean even within certain cities where their accuracy is really good haven't been adopted is a little bit because of this accountability issue where like it seems like humans are actually willing to have more death and chaos as long as they can clearly point to whose fault it is what I really believe to be true though is there's a certain accuracy threshold where if met we will change that a certain benefit where if we really can experience this benefit we're willing to think in a new way.

16:50But it's not just enough that it's better. It has to be better. It has to be faster. It has to be more observable and traceable because you're right. Humans are already managing humans. And that's actually not that easy as I'm sure you know. And shortcut specifically, on a scale from

17:102005 self-driving car to 2025 self-driving car, where are we? Yeah, great question. Think of it in two dimensions. One is, or two almost, let's do two kinds of use cases. There's a use case of like build something from scratch. And let me show you, for example, this one, right? Bam. So this was the one I asked it to build something from scratch, or more or less. I said, here is, you know, data, build me, build me summaries, dashboards, right? And in this use case, it just did this in eight minutes, Greg. That would have taken me an hour or two, and I'm good at Excel. For use cases like this, we're already past this self-driving moment.

17:52We're past the way-mount moment. Now, and I'll continue to answer your question, but I'll show you why it looks like this, which is super cool. This is the observability we're talking about. You can see exactly what figures are hard-coded, which are formula-driven and why, so that you can really review this better. But the real use case, like I was pointing to earlier, is you're going to update, not things from scratch, but you're going to update existing models. Now for this use case, which I think is 90 % of real Excel work, where the billions of dollars are, I'd say we are at the Tesla self-driving right now.

18:26As in, if you're in the know, if you're willing to adopt frontier tech, it's very exciting and actually useful in your workflow. But humans are still, by and large, doing the manual driving themselves. I like to think of it as we're probably heading for an August 2024 moment which is when Carpathie tweeted about cursor and it went like 20x I think the question is is it August now or is it May? but there's a very clear line that we can future predict towards that looks like we know what to solve and it is solvable so it is, as you can imagine, super exciting for us and it's also very clear that a product like this should exist like using plain English to we've had Vibe coding, we've had Vibe marketing, we haven't had Vibe Excel yet but I think this idea around taking your thoughts and building in this case it's not code but it's kind of quasi-code in some ways right?

19:33Yeah of course I wish I could show you the logs because it's, of course, code. Everything is code, right? Right, and everything will be code. And I think all code will be generated dynamically. But the thing you're getting to is what I've always thought was, maybe one of my smarter ideas, but it's not mine exclusively, is the best ideas are not definitionally contrarian. As in, I didn't have to think of something that you didn't believe and then make it true. I just think the best ideas are really obvious in hindsight. And I think the best predictor of what that is, is if you can release something, do people say, I can't believe this didn't exist already?

20:08Right? So when people see this, I think that's the reason that the initial reception has been so strong. It's just, why hasn't this existed? Like, there has to be some explanation. And I'll tell you what it is. We're a research lab with 20 people that are all, you know, from MIT, Stanford, great researchers. But if it wasn't for my background in finance, we wouldn't have done this. And I just don't think that there are co-founders at the 10 or so frontier research companies in the world that care at all about finance. No researchers really spending material time in Excel at all. They are building coding agents because that's what they love.

20:45So it took someone of a little bit of a different background to make it happen. And now I think the world knows how important it is because of the initial reception. And I'm sure everyone's sprinting towards it. Yeah, I mean, that's why I reached out. because I had this in my brain that I wanted this. And so it's cool to see it working. So what's happening? What am I seeing on screen right now? Yeah, so this is actually a great one. This is the big hard task, right? Which is update the Microsoft DCF model using Google's new data. Just use the same exact template, don't change anything except for the data.

21:23And what you found is the agent has actually extracted all of this data, has updated the historical data, the drivers and the assumptions, is looking at the tax rate and it's actually kind of finding errors as it goes. And it's like, that actually doesn't really quite check out. I think CapEx is too high, right? Here's an example. There seems to be an income calculation error. Let me correct the issue. I see the issue. It's that in the data sheet, it has the wrong references and it's always finding its own mistakes and just chugging along. So you're seeing it actually do the kind of work that a human would have to do.

21:53And it's about, I see the issue. There's a dependency see and it's causing a circular reference. One of the hardest problems in Excel, you have these things that are codependent on each other. And it found the circular reference in the exact formula that it was referencing, which isn't a hard formula. I don't know if you ever use sums in Excel. But it found it. And it found it, and it fixed the error. So that error is gone. And now it found that there's a couple ref errors in this SGNA. And it's correcting those two. And now that's correct, and there's still some remaining ones. And then Greg, what I'd also like to do is, if there's anything you want to try, go push it.

22:30Let's go break this thing. I'll tell you one thing that's been on my mind recently.

22:39We've got an agency, a design agency for AI companies called LCA. One of the things that if you run a tight agency, you need to have utilization rates you need to have people basically on files they need to be utilized because the big problem that a lot of agencies have is at the end of the year they're making 5 % EBITDA 3 % EBITDA, 7 % EBITDA they're very hard businesses to run because people aren't utilized so I wonder if there's a way to create some sort of profitability analysis around agency utilization rates based on different teams and stuff like that. Ultimately, my dream, Nico, is to look at a spreadsheet where it says, this is how utilized this team is, this is how utilized that team is, this is how much revenue they're bringing in, this is how profitable they are.

23:49just like a bird's eye view of utilization and profitability. You can tell my spelling has deteriorated ever since using AI. So let's ask Shortcut for that. It'll ask clarifying questions and then we will see what we can do about it. Okay, cool. I mean, if this works... This is a good test. This is a good test. I had done something similar where I asked it like, hey we're going to do some kind of launch soon I'm looking for a bunch of the most in distribution best launch partners that I can meet with, talk to and I had to find that across every single platform and then their posting schedule why it's a good fit, how much they expect it to cost even they got their people's emails so it's kind of a similar kind of task like that so you tell me this is your dream task right?

24:39do you want this multiple sheets? custom? single sheet single sheet what time period should the utilization analyst cover monthly key metrics yeah utilization percentage revenue that's great cool all of them do you want a dashboard to visualize the data yeah why not yeah make it hard yeah just make it hard there's now like what kind of data it's going to assume to make dummy data I suppose you want me to like get weird data from the web like what do you want to do is there a way to benchmark our utilization rates versus standard design agency utilization rates? I want to know if we're ahead of the pack or behind.

25:28Maybe I'm making this way too complicated for you. No, let's do it. Can you benchmark existing and standard agencies? These are which ones specifically? Design? Design agencies. And then for our data, Do you want me to say, use Greg Eisenberg's company, see what you can find? Yeah, I mean, you can go latecheckout.agency. Go to latecheckout.agency. That's my company, sure. Yeah, so I mean, it's pretty vague. Let's see if we can start marching towards your dream. That's all I ask, because I I feel like I bug the finance people on my team, and they don't want to hear from me. Like, add this column here and do this there.

26:21And I also think that a lot of people listening are like solopreneurs, are small teams, startups who don't have finance teams. Right, right, right. Right? So is this your finance team in a box type thing? Yeah, I mean, that's why I used that first prompt here, which was like, I have my expenses. Let me walk you through the actual answer here. I don't have a finance team. I'm one person out of 20 and I'm the only business-oriented person and I still spend most of my time coding. So I had the expenses here. I actually need to build a P &L. So pull this up. And in fact, actually there are some errors here which I'm not used to seeing.

26:57But let's go into why. Historical P &L projections. Oh, interesting. So it projected the revenue out for the fiscal years according to where these sources of revenue were in the expenses. And it made a dashboard here. So some bizarre formatting choices, quite honestly. But you'll see it's looking like it has revenue. And these are all formula-driven, so you can see exactly where they come from. Some of the ones are going to be more assumptions. You see total revenue, gross profit, operating margin, and so on. Even charts it. and it took my data and built me a forward-looking P &L. Crazy. I think we're probably, for these net news, it's like having your own mini finance team.

27:42Because you don't want to do this from scratch, but you probably would want to just do your last tweaks at the end. Yeah, exactly. So here, actually, we're going. You ready? Yeah. Let's see. Sheet 1 exists. It's going to create a profitability analysis for a late checkout agency. Well, interesting. By the way, Greg, I've never seen utilization sheet like this. So tell me your thoughts as you're seeing it. So, I mean, going left to right. So obviously I love the breakdown of employee and level. like that's we level people and you know higher levels obviously get paid more and stuff like that so i i love that um available hours by month it's interesting to see it like that but that's not exactly how i envisioned it like to me i'm kind of like a percentage guy like do you have does jane have 10 next month okay let's go you know maybe let's not put her on a project because we don't want to be at 100%.