Excel AI capabilities in Claude and Microsoft Copilot explained with prompt examples
Nate B Jones of AI News & Strategy Daily walks through how to use AI — specifically Claude and Microsoft Copilot — to automate complex Excel work, with four detailed prompt examples.
Summary
Nate B Jones argues that the recent launch of Excel capabilities in Claude (particularly with the Sonnet 4.5 model) and Microsoft Copilot's new agentic Excel builder represent a step-change in what AI can do — moving from language generation into production-grade numeric computation. He contends that AI for Excel delivers a far greater order-of-magnitude time savings than AI for words, because spreadsheet work is inherently complex, formula-driven, and multi-sheet. He walks through four specific prompt structures — a monthly P&L builder, a board-ready financial package, a three-year business plan model, and a data-editing workflow — explaining how each is designed to avoid context-window failures and handle ambiguous data. He closes by framing this moment as comparable in significance to the launch of the pocket calculator, and challenges viewers to invest the time in learning these prompts given the potential to save dozens or hundreds of hours per month.
Key Takeaways
FULL TRANSCRIPT
Why AI for Excel Is Different From AI for Words
Nate B Jones: If you are not using AI for Excel, you are sleeping on the power of what it can do. It is one of the biggest changes I have seen in the last year for AI. This video is all about AI and Excel. It's about Claude. It's about Copilot. It's about what we can do that we couldn't do before. And I go farther than that — I'm going to talk about specific prompts, show you specific prompts, and how I use them to save hours of work. I'm also going to give you a sense of how we should think about these new emerging capabilities in AI and how they change our own calculations of what's worth doing for people versus what's worth doing for AI, and how much we should pay for AI and invest time in AI when it continues to gain capabilities like this. Lot to cover. Let's dive in.
What Launched and Why This Video Exists Now
First, what launched and why I'm doing this video now. Claude has launched Excel capabilities over the last couple of weeks. It's extraordinary. It's really good. It got better with the launch this week of Sonnet 4.5. I was able to one-shot complex Excel analysis out of a screenshot for multiple currencies — like four or five different currencies — this week with Sonnet 4.5. It's gotten really, really good. But people don't know about it because it's in Claude's Max plan, which is expensive. We're going to get to that in the ROI part and how you think about that, so just put a pin in that for now.
If you're not on the Max plan, this is still for you. One, Claude is bringing those Excel capabilities down-market — you can look for them to come down soon. Two, Copilot also launched this week an agentic Excel builder. It can be in Excel, it can be your assistant, it can build Excel for you, it can edit Excel for you. I've played around with it. To me, it reads as about a number two compared to what Claude can do with a great prompt. But it's the tool you've got if you're in Copilot, and it's fantastic and you shouldn't sleep on it.
The larger story is clear. We had LLMs doing words for a long time, and now they do numbers. And that means you need to think about ROI and value differently. It also means you should assume that this video matters for you regardless of what plan you're in today, because the capabilities are coming to you. They're coming to you in three months. Everybody's going to have this — maybe in a month, everybody's going to have this. And frankly, if it's in Copilot, most people have it now already. That is how quickly this is changing.
How to Think About the Value of Excel Work
So how do you think about the value of Excel work? I feel like it's very different from words. When we are prompting with words, primarily what we are doing is we have a thought and we are saving ourselves the hands-on-keys time to write it all out. Whether you do that well or badly, that's the idea. You can prompt badly and get malformed thoughts or thoughts that feel like AI slop that aren't really your thoughts, or you can prompt really well and get stuff that is your thoughts just written out really, really fast. It's an accelerator. That's fundamentally what words are.
But it's different with Excel, and I think the leverage is in some ways higher, because what you get with Excel is the leverage to compute numbers. With Claude and with the new Copilot launch, that leverage is essentially production grade. It is not as good as a CFO who has been doing spreadsheets for 25 years. I don't want to overstate it. It is not as good as the person who has maintained the 20-tab sheet for 15 years in your office. No. But it is an extremely good professional assistant that can prepare dozens of different financial statements, models, attribution analyses, or ROI calculations. It can do all of the standard business spreadsheet work 30 to 40 times faster than a human, in a better-organized way.
I want to call out Claude in particular here. Claude is a good notetaker. Claude is so obsessive about having a little tab that has all of your notes there, and I love that so much. It's going to do that faster than anybody in the office, including the CFO.
When you think about the ROI for these investments — is it worth going to Claude Max? — I don't really think about it as am I going to get the money back per se. I think about it as am I going to get the time back, and do I need it now or can I wait a month for it to come down to the right plan. There is no question I will get value back in terms of what it can do for me if I know how to prompt it. And that's exactly what we're doing here — we're going to talk about how you prompt it so you get the value.
The prompts you are able to build — the revenue analysis, the board-ready models — these are going to save dozens of hours for finance teams. They'll save dozens of hours for PM teams who are trying to build ROI models. Excel is extremely hard for humans to build because it's all predicated on the idea that you have to have control over every individual cell, know what it says, and know how the formulas compute. That is inherently a very high-compute task for us. It goes in our brains. This is why I lose my hair. I had to do a 20-sheet Excel when I was a marketer back donkeys years ago, and the gray hair is real.
You have to think of these as multiplied accelerators when it comes to Excel. If LLMs just accelerate linear production of words, they're already amazing. But with Excel, they are giving you not just quick thoughts — they are giving you full models that you can use to do real work with. To me, that is actually more time leverage. It saves you another order of magnitude more time. If I am trying to write a product requirements document and I can get it done in 20 minutes instead of 3 hours, that's great. If I am trying to build a set of board-level analyses and I can get that done in 60 minutes instead of 8 days with a team of three, that's another order of magnitude of improvement. It's a big deal.
Now, I do want to emphasize that the entire workflow shifts and moves the data collection burden to you early in the process. You still have to collect the data for Excel regardless — to get all of this work done, that doesn't change. Same data, got to get it. You're building an analysis, you need the same input numbers, they've got to be right. But you have to get them up front now. For people who are used to saying, "Okay, I'm going to go to the next sheet, going to go get a cup of coffee, going to go pick this up off of Daryl's desk," and then come back and look at it — that's different now. You've got to get it all in advance, make sure you have it all.
Now, you can give relatively messy inputs. As I said, I gave a screenshot with numbers in multiple currencies and I got frankly surprisingly good outputs from Sonnet. It was like one mistake and I corrected it quickly, and that was it. I think we are sleeping on the leverage we get if we are willing to do just a little bit more work up front. So do the little bit of work up front, collect the data, and without further ado, let me show you four examples of prompts that I am building and using for AI in Excel and why I think they're so powerful.
I actually built a bunch more of these — I think I have like 18 of them — and I'm putting them all together into a prompt pack. They'll be in the post. But I want you to at least get a sense of the structure, because my goal is for you to understand how I put them together. There's not magic to it. It's really about communicating intent.
Prompt One: Monthly P&L From Raw Data
Okay, this is actually a relatively short one. These are in ascending order of complexity — I thought I'd start you out easy. This is not multiple hundreds of lines. This is just a few dozen lines per step. Our goal is to get a monthly P&L from raw data. Let's say you have a CSV export from your accounting system. It's a little bit of a mess, but let's put it together.
My recommendation is that you actually do these in pieces. You'll notice that these Excel prompts do come in pieces. They presume that you will get a deliverable at the end of each step. If you notice at the end of this initial step, it says: return the clean data in a new sheet called "clean data." Just in case, for example, you run out of context window on one of the larger prompts, this allows you to continue your work. It doesn't become one of those nightmares where the thing runs out of context and — as Claude at least does in a very frustrating way — just says, "I can't do this, we're out of context," and then you lose like half an hour. It's designed to prevent that.
So it's very simple. You declare your columns and then you declare your intent. "Please create a clean data table," and then you start to define cleanliness. Dates must be formatted properly. Amounts must be formatted as currency. Remove duplicate transactions. This is an example of something where, if I knew what was in the data set, I would probably be much more declarative about what counts as a duplicate to avoid false positives — you can expand that, make it yours. Add a column for month — that's something that would be relevant for most people to be able to filter. And then flag rows with missing data.
You can identify other steps in your data hygiene process. Most organizations will have a data hygiene process that they can go through — it's an SOP, a standard operating procedure. You can literally encode that as a prompt. It doesn't have to be a manual thing that you do every time you open up the sheet.
Then second step, new prompt. You use the sheet and now you start to do work against it. Now you basically take it and start to categorize it, now that the data is clean. You're going to define categories — these would be up to you. I made them up, but you can define what you want them to be. And then use the description. This is something only LLMs can do — this just was not possible before with rules-based categorization. I know I've written those sort of if-thens and they're nightmares. This is much, much easier.
The examples are good. If you feel like you have some areas of concern — like, I don't know if this is correct or incorrect, or a duplicate or not — you can include counter-examples. You can include tiebreakers or rules around ambiguity. And I love this at the end: mark that it needs review and add a note explaining why it's ambiguous. That makes a lot of sense because you don't want it to try and declare something ambiguous on its own — you want to give it back to a human.
Then we get to the monthly P&L. You create a new sheet. Again, you're declaring your rows, you're telling it what formula to use. One of the breakthroughs that Claude launched — and Copilot with Excel does this too — is we have fluency with formulas finally. And you can format it with bold headers, subtotal rows. Can you tell I've been having fun since Claude launched this Excel thing? That was the beginning for me. I was like, "Finally, it can do Excel. How much can it do?" Well, you're starting to see how much it can do.
Then we get to budget variance. You can actually start to compute budget variance here. You'll notice at each step we're just saying this is a new sheet, this is a new sheet. It will come back with a new sheet. And even if it runs out of context window, you're okay. Now you're doing one operation on it — just add the columns.
By the way, Copilot is really, really good at this — just do this one thing in the sheet. It lives in that embedded sheet, whereas Claude treats it like an object and manipulates it from the outside. And actually Claude is still very, very good. I think it's actually better overall in my experience than Copilot with Excel in the sense that it has more ability to follow complex long-running prompts and get high-quality results. But if you want the inline edits — the ability of the Excel magical AI — that's just something Microsoft has a distribution advantage on and they're just going to do it. So this is an example of something where I privately think Copilot is going to do well.
And then you can validate it. You can add a validation section at the bottom and make it prove itself and display it as a checklist. And this is what you get out of it.
I do have more — I have a sales pipeline dashboard and others — but this should give you an idea, if you were an individual contributor, of what you can do with Excel. You might think, "Well, Nate, how often do you touch monthly P&L?" You have to touch it every single month. And let's say it takes you hours and hours a month, and now it takes you 5 to 15 minutes. That's ROI that pays itself back really, really fast.
Prompt Two: Board-Ready Financial Package
Okay, we're going to go through this a little bit quicker now that you know the idea. This one is for a board. Let's say you're a founder, you're trying to put this together, you want your chief of staff to put this together — whatever the case. You need a comprehensive financial package.
You can go through and have it create a revenue analysis sheet with comprehensive revenue tracking. You're defining that all here, you're declaring it. These are very standard, and that's good — you actually want standard if you're talking to the board. It's just going to come back with the overall perspective. You're going to need to supply all the inputs for this, so you should add that in when you write this prompt. Then it's going to come back with the growth analysis, the calculated metrics, and all the rest of it.
Then you have a separate step for cash runway. You can go in and have inputs at the top for cash balance, for month, for burn rate. You get your projections, you get your alerts, your milestones, and you even get a visualization. And if you're wondering if it makes visualizations — yes, Claude in particular is very good at visualizations in a way that I am still surprised by. When I ask it to make a chart, it makes a good chart.
Then you can give it this little instruction to format for easy copy-paste into board slide decks, which is a nice touch. You can do all kinds of things people don't realize — I have told the AI where to bold, and it will just bold it in the right place. There's all kinds of stuff you can do that people don't realize, that make it seem more professional.
Step three, you want to get to a key metrics dashboard. At this point, honestly, you're just kind of dunking on the competition, because you don't have to have this as a separate dashboard — it can be — but I wanted to show you what you could do. So you can go through and calculate these metrics in a new sheet. You define how you talk about benchmarking here, you can define your metrics. Again, these are very standard metrics, so there's nothing too surprising, which is good because you actually want AI working off of standard metrics, and also because you want your board working off standard metrics.
Department spend breakdown — very similar. You want to go piece by piece. Are you getting the idea? This may be hundreds of lines in total, but we're breaking it out so that it doesn't break the context window. And it's going to save you days of work. It is worth figuring out how to use it so that you can save those days.
There's a ton here. Now you're formatting it. Here's a board summary sheet. You talk about what's in the overall report, you give it very specific formatting specs and export steps. And honestly, this is very Claude-specific, because Claude can work across PowerPoint and PDF and Excel within one prompt, whereas Microsoft's Excel agent won't. So I wrote this for the more complicated use case, but you could obviously drop it out if you'd like.
The 13-week cash flow model is a totally different thing. You get the idea — it's a longer prompt, broken out in ways that work for the context window, uses standard terminology which is good for the LLM and the board, and presumes that you can gather the inputs you need.
Prompt Three: Three-Year Business Plan
Let's go up yet another level. Let's say you want to build a three-year business plan. Now, you might think only founders need to do that. But honestly, everybody should build a business plan if they're doing anything at all related to a side gig. And everyone should be able to read one if you are in a business, because boy does it shape your world whether you like it or not. At least you should be able to understand this.
And by the way, one of the things that's really interesting about AI as a sidebar is that the availability of material like this makes AI a phenomenal teaching tool. You can go into Claude, ChatGPT, any AI, and say, "Please tell me why the three-year business plan has the elements it does," and you can get a whole lesson. You just get a whole lesson on it.
Anyway, we create an assumptions sheet which serves as a source of truth. We start to do that first. We go through each of these. You'll notice we are getting very clear with our seasonal patterns. You will need to go through and hand-adjust these so that you have the right adjustments for your business. Similarly with hires — you want to be able to hand-adjust this. We get to expenses, et cetera.
Now, the good news is, let's say you were like, "Oh, Nate, I don't want to go through and touch every single line." Well, you don't have to. You can go in and copy this prompt, then get all of the spreadsheet — the dirty CSV, the picture of the cocktail napkin where you have the data — and say, "Put A and B together in a smart way. Here's my prompt. Do not follow the prompt literally. Instead, fill in the prompt with this information." And the LLM can do it. Any cutting-edge AI will be able to do that.
You go through and say, "These are all of my assumptions." You go down to step two and now you're building a revenue model. It's a separate sheet. You're going to create the revenue model and stack the cohorts all the way through. Make sure that it understands how you want to build your business. You'll need to adjust this for your business. We're assuming implicitly almost a B2B format here, because that's become the standard for a lot of the way we talk about business in the software world. If you're not an MRR business, don't use MRR — use the metrics that make sense for you.
Then you go through and create the visualization. The point is to show you what you are capable of doing in these models. Then step three, you get to expense models by department — another level deeper versus the board-ready package. I don't want to bore you. This is a long prompt. The good news is I wrote it already for you. All you have to do is modify it.
You can do the headcount model. If you're a tiny startup, you don't need that — that's part of why I broke this up. It's all optional. You can do a P&L rollup — again, it's more complex than the board one, but it's good for a business plan. And you can do finally a cash flow statement. There's a lot there, but that's good, because at the end of the day you actually get your entire business plan done. This is the kind of thing a bank needs, or a venture capital firm needs to think about investing. It's much more in-depth. But this kind of thing saves you hundreds of hours if you do it right.
Prompt Four: Editing and Cleaning Existing Spreadsheets
Let's look at one more. And I saved the best for last. I saved the edits for last. Yes — I want to talk about the possibility of editing with Excel. That has been really challenging for models in the past, and I am so excited to share what I've worked on here.
It is really, really important when you're editing to be clear about what you want. So this is going to presume certain kinds of mess. My goal here is to show you how to clean up as a set of principles, so you can clean up your own mess with prompts that are forked from this.
For example: "I have a messy data export from —" and then you'll have to specify where it's from. "This is the current state." You want to describe this as specifically as you can. So: inconsistent formatting, duplicate entries, text in number fields, unclear column headers. And then: "This is what I want." And then specifically, "This is how you clean the data." You want to be as specific as you can here, and it should map back to the current state.
This is what categorization looks like if that matters to you when you're cleaning the data — again, you're going to specify that as cleanly as you can. This is how you start to work against it. And then these are your quality checks. The output is a clean data set with a summary dashboard showing total items by category, monthly trends, and items requiring review. Ensure all formulas are error-free and totals reconcile to source data.
You can actually do more than this in validation. If you get paranoid, you can make it produce a separate document showing how it did all the work, so it can go through and explain to you how it did it as a way of auditing.
What would you change with your mess? That is my question for you. Would you change the data-has-issues section so it's more specific? Would you maybe say, "80% of our issues are inconsistent formatting, and these are the six ways we see it"? Those are all ways to give the LLM more clarity. Basically, think of this as the LLM's briefing before going into the spreadsheet. You want it to be as comprehensive as you can so the LLM gets no surprises and doesn't hit a point of ambiguity.
If you think it's going to hit a point where something might have inconsistent formatting or a mixed date format and there might be two correct answers based on your prompt, your prompt is not clear enough. And that is one of the reasons why edits fail.
I've come through and done some other edits — like you can get it to fix broken formulas and all of that. These are complex prompts, and these ones, because they're edits, are going to require another level of involvement from you, because everyone's spreadsheet is broken in its own special way. You will have to go in and talk about how your spreadsheet is especially broken to get this stuff fixed.
But this is really, as far as I know, the first real effort anyone has made to make AI do edits and not just create from scratch. It is harder because models are trained to create from scratch — that's how reinforcement learning works. It is still worth it to figure out the edit piece, because if you can get a cleaner version of your spreadsheet, if you can fix your spreadsheets, again, dozens of hours saved. It's a huge time savings.
The Bigger Picture: What This Moment Means
Okay, you have been patient enough to stare at four different kinds of prompts. You've looked through IC-level prompts. You've looked through executive-ready board meeting prompts. You've looked through business plan prompts. You've looked through how to edit with Excel, which is net new.
I hope what you walk away from is realizing that you have the ability in your pocket now to save hundreds of hours on complex business computation. To me, what has launched in the last two weeks has the same leverage and power for numbers as the launch of the pocket calculator. That's how big a deal it is. And we're sleeping on it because most people either don't have the Claude Max plan, haven't really sat down and thought about prompting the Copilot AI for Excel, or they just don't have the time to come up with these complicated prompts and they've always done it this way and so they're just always going to do it.
Let me challenge you. It is worth it to go a little bit differently, to try a new prompt, to adjust one of my prompts. If it saves you that much time — if it saves you a dozen hours this month — the Claude Max plan is probably worth the money, because you're going to spend that amount of time regardless. Your time has some value to it.
I continue to come back to this idea that AI for words was huge — it was like a 10 or 100x speed-up. But the potential with Excel is a thousand-x, because of how complex these numeric structures are. They're not lists of numbers. They're complicated multi-sheet, hundreds-of-formula, multi-cellular structure messes. And now we can finally deal with them.
This has been, I really think, the biggest week for leveraging numbers and going farther in business since the personal calculator. It's huge. Maybe since the original Excel — the original Excel was pretty cool, maybe we'll give them credit there. But it's a big deal. So have fun. Go dig into Excel.