Claude AI Replaces ChatGPT and Copilot in Finance to Automate Workflows without Planning Tools
In this episode of Future Finance, Paul Barnhurst and Glenn Hopper explore how AI tools like Claude, Copilot, and Excel agents are transforming financial workflows in 2026. As the finance industry continues to evolve, AI is playing an increasingly crucial role in automating routine tasks, improving accuracy, and boosting efficiency.
Paul and Glenn discuss their firsthand experiences with AI-driven tools, from tracking expenses and managing journal entries to building financial models. They also dive into the ways small and medium businesses are utilizing AI-powered Excel agents to streamline their financial processes without the need for expensive, traditional planning tools.
In this episode, you will discover:
The impact of AI agents in finance and how they simplify workflows.
Personal experiences with Claude and Copilot in real-world finance tasks.
How to integrate AI into your finance team’s daily tasks.
The benefits and challenges of AI-powered automation in finance.
Paul and Glenn emphasize the importance of embracing AI in finance, highlighting how tools like Claude, Copilot, and Excel agents are transforming everyday workflows. They encourage finance professionals to experiment with AI, even if it means starting small.
Follow Glenn:
LinkedIn: https://www.linkedin.com/in/gbhopperiii
Follow Paul:
LinkedIn - https://www.linkedin.com/in/thefpandaguy
Follow QFlow.AI:
Website - https://bit.ly/4i1Ekjg
Future Finance is sponsored by QFlow.ai, the strategic finance platform solving the toughest part of planning and analysis: B2B revenue. Align sales, marketing, and finance, speed up decision-making, and lock in accountability with QFlow.ai.
Stay tuned for a deeper understanding of how AI is shaping the future of finance and what it means for businesses and individuals alike.
In Today’s Episode:
[03:47] – AI Agents in Excel
[06:12] – Claude vs. Copilot
[09:27] – Automating Expenses & Journal Entries
[15:56] – AI for Financial Models
[21:21] – Future of AI in Small Business Finance
[26:26] – Workflow Automation in Finance
[35:28] – AI & Human Collaboration
[39:31] – Experimenting with AI in Finance
[41:45] – Data Quality & Governance
[43:07] – Key Takeaway
Full Show Transcript:
Host: Paul Barnhurst : Welcome to another episode of Future Finance. I am one of your co-host Mr. Ebitda. Oh, wait, Paul Barnhurst, the FP&A guy, and I'm lucky enough to be joined again by my partner in crime. Glenn, how are you doing, Glenn?
Co-Host: Glenn Hopper(01:30):
I'm good, man. I was going to say we've been missing each other this week. You've been emailing me and I haven't been responding. I'm like the ex-girlfriend that's making things weird, but I've been head down on a bunch of work, but making my life better with AI every day. So I guess it beats the alternative to not having anything to do with
Host: Paul Barnhurst (01:47):
It. Yeah, I've been realising more and more how much I need to spend more time on AI as I keep digging into it, building some actual agents, doing more, but I've also spent a fair amount of time with Excel's agents, so it's fun. So you mentioned it's making your life better. I know when you and I chatted, you've mentioned you've been making a bunch of changes. So I think for this episode, would love to first hear the changes you've made and some of these breakthroughs you had. Maybe we could talk about how they relate to finance and we'll go from there.
Co-Host: Glenn Hopper(02:16):
So first off, I wasn't trying to make a controversial post. I wrote in my substack, it's been several weeks now. Maybe I put it somewhere, I put it in my substack or LinkedIn, and I got a lot of flack for this, but I stand by it and even more so I was an early chat GPT user and I thought I was committed to it because I loved its memory and it knew me and it was tuned into my writing style and it was just a good assistant across all the projects that I'm in. It was able to pull in context, but it seems, and this isn't, oh, I miss four oh and the way the Sycophant way it talked to me or whatever. This is even more recently from five one to five two, and then I don't know what they're testing out now ahead of five five, it's going to come out, but Oh, is five
Host: Paul Barnhurst (03:06):
Next? Can I say something? Do they know how to number?
Co-Host: Glenn Hopper(03:09):
Yeah. Well, the thing is, in version control in software that's normal to have the point, whatever releases if it's not a significant release. But now because their product naming is so bad, this isn't just open ai, this is across the board, all their product naming is so bad that we are all now in tune with a software version numbering, which we shouldn't be. It should just be this is whatever. I've been pretty frustrated with chat GPT hallucinating forgetting rules, I've given it, and at the same time my claw use has gone up significantly, especially around modelling and anything where I have real finance and accounting work to do and what we're using on the backend for some of these workflows that we're building for clients. The straw that broke the camel's back was some bad interactions with chat GPT, where it just was not helpful.
(04:05):
I was trying to hand off or bring a new resource onto a client where had all the client info in there and it just was spitting out this, I was trying to create a document where I was handing something over to a new, very seasoned person and it was talking like it was a kindergartner coming in and I kept telling it, change your tone, write this as if this were an internal memo at McKinsey or whatever, and it just was terrible and all these weird insulting parentheticals. And I was like, why would you put that? I didn't tell you to? Anyway, so frustration with that. Meanwhile, I should be embarrassed to say this, Paul, I've been in finance my entire career, but I'm sort of that what's the crying clown or the cobbler that needs new shoes? My personal LLC I've had for a decade, but it's always been, that's just what I run side hustle gigs through and I've been really lax on tracking expenses around it and overpaying the government for years.
(05:02):
But this year with all my speaking engagements, my training, my webinars and all that, I ran that through the LLC rather than through primary business. It was higher than my W2 income, and when I realised that on December 4th, I thought, oh, I need to offset this with some expenses. So I had two credit cards. This is a really long story, I'll try to condense it down, but I had two credit cards that were co-mingled. I had personal and business expenses on both. One was primarily personal and one was primarily business, but neither were clean and the one
Host: Paul Barnhurst (05:36):
Weren't a typical small business, you just weren't running a contribution through revenue
Co-Host: Glenn Hopper(05:42):
Owner
Host: Paul Barnhurst (05:42):
Contributions. Good. Yeah.
Co-Host: Glenn Hopper(05:44):
So I thought, and I haven't talked to my bookkeeper, I love her, but I haven't talked to her in 18 months and I thought she's going to kill me when I give all this data with no context. So trying to make this as short as possible, but I want to give enough detail to express just how valuable Claude was in this situation. So thousands of transactions on the personal card primarily, and then my business card, fewer transactions, mostly business, but we were going to have to go through, I say we, me and my wife, who since I haven't talked to my bookkeeper, we're going to do this. We're going to have to go through every month of statements and pull out these are business and all that. So I did some context prepping and I said, look, this is the travel I did last year, which I got by the way, by having Claude go look at my Google calendar and pull all my trips out.
(06:35):
And then it identified, okay, this is when you did travel. And then I said, go through and look at webinars. And then it pulled out when I was hosting anyway, so I had all these various associated things that Claude pulled out for me and I said, now go look through these. And I was actually able to get it down to just CSVs of every transaction. So go look through each of these. I did one card at a time and it said, identify all possible business expenses, pull them out on another CSV. Here's a hint. Things that are happening in other cities around these dates will be likely most likely business. But it even pulled out things like I was paying on my personal card, my Google workspace account. I was like, I bet that's for business. And I was like, yeah, that is, that's for, so anyway, I identified all those that went through the other card and did the same thing, got a good number of files or a good number of transactions, did the human in the loop.
(07:27):
I went through all of them, made sure it wasn't putting my YouTube premium subscription in business or something like that. And then I did project accounting. I went through and I looked at, okay, these are my trips, and I had it tied expenses to every trip that's not for taxes, that was just for me. But then I uploaded my chart of accounts, and by the way, I did all this, the first part of it, a big chunk of it with the Excel plugin for Claude, which we'll talk more about plugins after I get done with my ran. So I had it categorised 'em by business, and then I had some other expenses. And so then I had the two CSVs from the different cards and I said, okay, now I haven't paid these out of the business, but I had owner distributions. I need to offset the owner distributions.
(08:15):
Can you recommend journal entries for me? So I did quarterly journal entries and my idea was, okay, well now I'm just going to call my bookkeeper and say just make these entries. I know it's four entries, but it's been a while since I've done bookkeeping and I didn't trust myself, so I was just going to give it to the bookkeeper and have her do those. But then I thought, well, I wonder if there's an MCP for QuickBooks, which there is, but it's in a sandbox environment and it only came out in October and you can't do anything with it. But then we've talked before, I've been using Claude Cowork a lot, and I said, Hey, can you do that through the browser control? Can you make journal entries to me? And it was like, heck yeah, I can. And so basically in one afternoon I'd been worried about this really from the entire fourth quarter of 25 and then into getting ready to get everything over to my tax accountant.
(09:08):
And it just in one afternoon did all this for me and I stayed in the loop on it and I shouldn't be so surprised because we build things for large companies and I talk to individuals and companies about how they can use it, but when it's that impactful to you, I thought this is amazing. And what I just saved and bookkeeping and everything I would add to do around that, that would pay for a good chunk of upgrading to the quad max for the year. So after my frustration with chat GPT, I went down from my max subscription that down to the teams account that I have or whatever they call the business plan.
Host: Paul Barnhurst (09:49):
So like $20 a month one, the basic
Co-Host: Glenn Hopper(09:51):
Plan.
Host: Paul Barnhurst (09:52):
Yeah, that's what I have.
Co-Host: Glenn Hopper(09:53):
And then I upgraded Claude and oh, and it was interesting, and I probably should do a post on this, how do you extract all the meaningful data from one LLM that it has in your memory and bring it over to the other? I didn't export every conversation, I don't need every conversation, but I did pull what was in memory and I pulled my projects over and everything, and I'll be saying this, and then in three months I may say, well, now I've switched to Gemini Ultra because these models are always beating each other. But I think the writing's been on the wall for a while. Going back to my original statement that I think chat GPT, my wife and kids love it. I think it's more of a consumer grade. I think Claude is making real inroads into businesses and that's reflected in the market.
(10:39):
I dunno if you saw, I guess when this airs, this'll be last week, all the SaaS companies, stock software companies, stocks took a hit because Claude Cowork is automating so much of what they do. And I know before I went on my long rent, we were talking about one thing we could talk about today is automating workflows and how a normal person would go through and do that. But I think after my ran, we can save that for another session because what I've really wanted, and also we've got a mailbag thing, but before we get into our mailbag, because a couple of these questions were around this, I'm fired up about Claude in Excel because it's so much better than what I used before. However, you've been in copilot a lot and I know I've dogged copilot over time, but you're seeing improvements and I'd like to talk about that for a few minutes and then maybe we dive into our mailbag and we can, this stuff moves
Host: Paul Barnhurst (11:33):
So fast and then really testing it for real use cases versus just, Hey, build me a model. And you're like, okay, so what I will say a couple things really interesting this week. I'll start with this. So I'm working on my data visualisation course right now and I have an AI section and I decided to take some fake data. I'd taken a p and l, then I'd normalised it into a table, generally a little easier to read, although they're getting so good at reading the regular p and l that I'm not sure it makes a huge difference as long as it's a clean p and l, whether you normalise it or not. And so I'd done that and I decided to run it through an agent and asked it to create some graphs. I ran it through copilot. I first told it, analyse this data and give the recommended graphs.
(12:17):
It went ahead and built them copilot even though I didn't directly ask, but they were decent, they weren't great. There were some colour issues and didn't follow all the best practise. I probably could have prompted it and got there, but they came up with some graphs that I call pretty good. So I gave the same exercise to Claude and I look at it and I got all done and it had this chart, the words chart in three different places and it brought over data and the data was just a table. And so I said, where's the chart? And they're like, it's right there. I'm like, well, I want a graph. These are supposed to be visuals. And it just spun and never gave me an actual graph, even though when I ran that same exercise through the LLM, it produced three graphs using JavaScript or Python or I can't remember what it used to write it, but using one of those codes.
(12:58):
So my first experience with Claude in Excel, even though I hear everybody raving about it, was like I got better results using copilot over here. At least attempted to build the visual. So that surprised me. Now take that as one limited use case. So I fully recognise that's not real big. But what's interesting with these agents, and I posted about this on LinkedIn, is a couple things then I'll share something. What we're seeing right now is a blurring, I expect small to medium businesses to turn more to excel in the future than traditional planning tools. Not to say traditional planning tools are going away. Not to say they're still not a big market for them and a need, but that smaller and medium are going to go to these hybrids where the tool is just built as an agent in Excel. Yes, there may be a database and there's some out there, but they're going to streamline and be lower cost.
(13:54):
What you're seeing with several of these tools is they're becoming workflow tools already. Ebitda, AI pulls in campfire, it connects to Xero tabs, AI connects to Xero and QuickBooks, and you're seeing some of these new analyst tools building agents with them. So yes, some of the planning tools will build their own Excel agents for sure, but if you're a small company, do you need the heavy, most of the music sell today anyway. So does that extend, let's say a company hits 50 million, they start looking at a plan until you've been there a hundred million whatever, you really start to grow. You deal with all the headache with agents and having an agentic layer and Excel. Can you get to 200? Can you get to two 50 before you need a tool? I think it's going to increase when people decide to switch. There's going to be higher numbers as these agents get better.
(14:51):
And then something that's really interesting that I want to share, so I'm going to bring something up on the screen and then we'll get to the mailbag. So I've been following this for a couple months now. The first version I saw three, four months ago I think had co-pilot, co-pilot sell. It was like Chachi PT agent at like 45%. And then shortcut, I think when it first came out was like 56%. This is a benchmark score by spreadsheet benchmark. Look at how much the agents have improved. Now all of a sudden we're up around 70% accuracy. One I saw had humans at like 72%. What's interesting is two of the top three are now coming from China Kings soft office, which is the Microsoft knockoff in China basically. And so I thought that was really interesting. VIR has already been tested and they're still in beta. They haven't even come out. They're not testing everybody. So it's rapidly improving. I mean to go from 50% basically a few months ago to 70 now, I mean that's a 40% improvement, right? 20 points, 50 to 70. So it's amazing how quick this is moving. I feel like anytime I say something, am I the curmudgeon? I'm questioning how good it is with how quick it's moving.
Co-Host: Glenn Hopper(16:08):
Yeah, that's interesting. And it's across the board seeing Chinese companies at two of the top three there. That's reflective of the broader what's happening in AI in general. And I think the difficulty right now is not to be an alarmist, however, it's sort of like quantum tech, whoever gets there first, it's significant in that you win and that it's a zero sum game here where there are consequences down the line if you don't win this. And I think, I don't know, there's a lot of bluster and talk about the Manhattan project for ai, and I don't know what at a state level is, but if you look at infrastructure and what's come before states have been involved, and you can't assume on any level that the US is just immediately going to win the AI race, and I don't know what that means, but that's a whole other road to go down.
(17:14):
But it's like legislation or didn't go out around social media and understanding of the internet and what it is and how it works. These are career politicians who don't understand technology, and I don't think they're being advised well enough on tech policy. And I know current administration is tech forward and crypto and somewhat on AI and all that, but I just don't think we are putting the right resources there. But that's a whole other tangent that maybe isn't even suitable for future finance. But it's worth noting that we are not the runaway and clear winners in the AI race globally.
Host: Paul Barnhurst (17:50):
It is interesting to know, right? It's important to remember, we don't know who the winners will be here. There's a lot of countries developing and how that impacts what we can use. Do we have access to the best technology? It plays a role. We won't get into it a lot because like you said, it's probably not the conversation we want for finance, but I think you had mentioned the mailbag. Maybe you want to ask me a couple questions from there. This is people that submit questions for a session I'm doing here and I thought it'd be really interesting. And then I'm going to ask you a few questions about what I'll call a agentic workflow. I know you're very particular about agents versus the Gentech workflow.
Co-Host: Glenn Hopper(18:26):
Well, and maybe your pedantic
Host: Paul Barnhurst (18:28):
As I like to say.
Co-Host: Glenn Hopper(18:29):
Yeah. And maybe the questions are just two sides of the same coin because first off, Microsoft, and this has been a pet peeve of mine when you're calling things agents that aren't truly agents, and I don't understand. I had talking to a client today about a billion dollars in revenue public company that they said they do most of their work in FP&A and Excel, which so your reference to small businesses, I mean huge enterprise
Host: Paul Barnhurst (18:56):
Companies, well even large businesses, even if you have a planning tool, some level of your forecasting, some level of work is happening in Excel, right? Excel's not going anywhere. What I think my point was more around can you get away with not having to have a special built planning tool longer especially, is more and more of the ERPs managed consolidation, is power BI and BI tools get better at the reporting or having an agentic layer, does that allow you to go longer and longer? Could you just take from what is Excel is and push it to your warehouse for reporting purposes of your forecast with AI with that layer? So that's kind of more what, I mean, there's things beyond just the planning, but that's kind of where I'm coming from.
Co-Host: Glenn Hopper(19:35):
And I saw someone in internal from Microsoft in their finance department talking about the agents that they were building and the automations that they're building internally at Microsoft. And I've just been so frustrated with what copilot can do in PowerPoint versus how is copilot not just integrated with design. Anyway, a whole other tangent, but I've just been like a lot of people
Host: Paul Barnhurst (20:00):
Frustrated with it's mean design suggestions. When you click on it, why doesn't it just use copilot?
Co-Host: Glenn Hopper(20:04):
Why doesn't copilot then? Even though the design suggestions sometimes can be terrible, like copilot, when you talk to it in PowerPoint, it's like, or even if you have it create a slide deck, there's no design, there's no anything to it, and it's like, you want me to drop this in your company template? It's like, no, I want you to drop it out of a window. It's terrible.
Host: Paul Barnhurst (20:22):
I will admit the PowerPoint. I don't ask it to help clean up my slides. I use it to create images, and that's mostly about it in PowerPoint, copilot office outlook. I'm using more and more as a search. I also will go into agent and say, go through all my emails for the last week and put me a summary of what I haven't responded to, the action items, the things I need to do or go through all my transcripts from teams and I can have this Excel list. So I've done things like that and those are pretty cool. So I think there's definitely benefit, but I think everybody, at least that I've interviewed has felt like copilot has moved slower and hasn't achieved its potential. And we see it in the revenue shortfalls. We've seen big predictions
Co-Host: Glenn Hopper(21:04):
Way
Host: Paul Barnhurst (21:05):
Short of what they anticipated. We all know it hasn't net forecast.
Co-Host: Glenn Hopper(21:08):
Okay, maybe Microsoft stock took a hit because of their just been behind on this, and I'm sure they're in a code red situation like OpenAI was a couple months ago with Gemini. And anyway,
Host: Paul Barnhurst (21:21):
I've heard they have deep pockets. Is that true?
Co-Host: Glenn Hopper(21:23):
Not at this rate of spend, right? Microsoft goes bankrupt on AI spend.
Host: Paul Barnhurst (21:30):
Well, you saw 11 labs was just raised 500 million on a billion valuation, or is it 10 billion? It was 10 billion.
Co-Host: Glenn Hopper(21:38):
Yeah, I think so, yeah. And
Host: Paul Barnhurst (21:39):
I was seeing another one where you saw, I'm sorry, we're getting it a little news. We'll get to the mailbox here in a minute, but I think this one covers ai. You saw Elon Musk combined his AI company, which has Twitter and SpaceX together, and the rumour is he wants to take it public later this year. Have you heard the valuation number? No.
Co-Host: Glenn Hopper(21:58):
1.5 trillion. So I'd like to see how much of that is made up from SpaceX versus
Host: Paul Barnhurst (22:04):
I think it rebelled at over 900 before the merger happened. What was brought in was supposed to be about 2 50, 200 50 billion. So yes, 75% of that is SpaceX. From what I'm seeing
Co-Host: Glenn Hopper(22:17):
And having spent a good portion of my career in m and a think about for SpaceX, there's zero value add of bringing them in. I mean, that is a way to shield XI in the massive losses, and Ty basically to me that I mean
Host: Paul Barnhurst (22:33):
Well, and to shield Twitter of not having to really show what's going on there.
Co-Host: Glenn Hopper(22:36):
Yeah, so I mean you can do that when you're, but if it were a public company and anyway, a whole other tangent to go down, but you would look immediately if you were a SpaceX to spin that off because it's a drag on. But if you want to bump up the valuation and your stock price and all that, on the same note, this gamble of what Musk is doing. So Tesla is no longer making whatever, two of the models that have been around forever. I mean, he's
Host: Paul Barnhurst (23:08):
Limiting a couple models. That was a big, I didn't look at which one. They're robotics. I saw the headlines on that.
Co-Host: Glenn Hopper(23:14):
Yeah, so they're going to be a robotics company now. So the things, if I were a shareholder in Tesla, which I wouldn't have the stomach for with all the meme up and down and valuation has nothing to do with business fundamentals, but I would be angry about that. It's like, okay, do your robot thing as a skunkworks side or whatever. Maybe it's the future, but we're not going to bet the existing revenue on it and hurt the business that it is as an automotive company, which I think Musk would say definitively, we're not an automotive company. So
Host: Paul Barnhurst (23:48):
He would, now, there's no question. He views it as a robotics company, which is name the last company that quickly, that big, that completely redefined a category that then all becomes a new category. And let's face it, whether you love him, you hate him, whatever, you think he's innovative, he's been incredibly successful from a dollars and cents in innovation standpoint. So who am I to bet against him? But it feels odd.
Ever feel like you go to market teams and finance speak different languages? This misalignment is a breeding ground for failure in pairing the predictive power of forecasts and delaying decisions that drive efficient growth. It's not for lack of trying, but getting all the data in one place doesn't mean you've gotten everyone on the same page. Meet QFlow.ai, the strategic finance platform purpose-built to solve the toughest part of planning and analysis of B2B revenue. Q flow quickly integrates key data from your go-to-market stack and accounting platform, then handles all the data prep and normalization. Under the hood, it automatically assembles your go-to-market stats, makes segmented scenario planning a breeze, and closes the planning loop. Create air-tight alignment, improve decision latency, and ensure accountability across the team.
Co-Host: Glenn Hopper(25:27):
It's whipsaw back and forth and all the moving, and it feels like a representative of the broader world of what's going on right now. So I don't know, this might be a sign that I'm just ageing and it's like, I don't know, I can't keep
Host: Paul Barnhurst (25:40):
Work. I hear you. Alright, let's jump into the mailbag. Let's pull out a few questions. You mentioned you had some for me and then I think some of these will dovetail nicely to just giving some advice for our finance audience on how to build agentic workflows. What should they be doing if they haven't really done any of that yet? Because that to me seems like that next level, okay, you're comfortable with an lm, you can do some prompts, but how do you start to get value beyond that and start automating tasks?
Co-Host: Glenn Hopper(26:09):
Yeah, and I think we'll see how deep we want to go in the mailbag. But first off, there are three or four questions here all about agents, and I don't want to go down another rant tangent on that, but because the questions are out there, I feel like you are the better person to answer these because you've had experience with it. So when people talk about agents and Microsoft, I have not seen them firsthand. I don't know what they're doing. But let me throw a couple questions at you and maybe lump them together and talk about this agent world within copilot. So lemme just give you a few of the questions. One, where do AI agents deliver the most? ROI and Excel today for FP&A teams? Two, how are CFOs and FPA leaders using Excel-based AI agents in real world finance workflows today? And three, and I think this will just hit these all broadly, how can FP&A teams practically build or deploy AI agents within Excel? So maybe for me, putting those together, first off, what even is an agent in Excel? Why are there different ones? How are people building their own? What are they doing with them? Why is agent different than just having copilot in Excel? I, I'm really clueless here.
Host: Paul Barnhurst (27:19):
So they're using the term agent and Microsoft's causing confusion as you know with that term. So the way I look at it is what it is, whether it's Excel's agent or co-pilots plugin or one of 33rd party tools or Gemini and Google. What all of them are is they're taking an LLM, they're building some kind of extraction tool on top of that, they're building some frontier modelling, whatever, and then they have a set of tools on the backend that LLM is going out to a tool. So the question comes in, build me this graph, okay? The best tool to build that graph and Excel is Java. The JS off a script for Excel that's in there or like trace light. They kick it off to a Python tool and bring it back as Python to you. And so the agents are basically taking in LLM, they have a set of tools.
(28:13):
And what's interesting, if you go into Claude and you ask it, it'll tell you all the tools it uses with its agent. I use that term loosely in Excel. Some of the others won't tell you, but you can ask it and it comes up with the list. Here's the read write ones I use, here's this and that. And so basically people aren't building their own agents in Excel. What it is is Microsoft have now coined the term vibe working and you can have that LL in there and you can select agent mode, which allows access to your spreadsheet and it directly works in your spreadsheet on your behalf. Kind of like Claude and what you did, you see how it can work in your spreadsheet. That's what Excel agents is. So that's the first thing. I think it's clear that people understand that most people aren't building their own agents in Excel.
(29:02):
These are third party companies. You pay part of copilot, you get it for 20 bucks. With Claude, you can get it, you can get a bunch of other tools anywhere from 10 to $500 depending on what version you're buying and who you're getting, what all they're able to do. And several of 'em are adding workflows to 'em. So some things that are really interesting with this, so ROI, it's early. I'd say most companies aren't yet getting ROI in FP&A investment banking. Several are using them. One of the best areas you can get is first great for helping with formulas, getting better at building things, but again, it's AI assisted. I've said this many times, those who know Excel well will build the best things. So for example, today I was building a deferred revenue schedule. I've done that exercise several times before, but I want to see how they did it.
(29:51):
Now I've improved my prompt. So that's part of it I'm sure, but it's a few months later, much more consistent between the two different versions. Much cleaner, simpler, even though they look, there's a lot of similarities between 'em. It's getting much better. And so that ROI at first is coming from help me build formulas, then other things, it's coming from cleaning up data. Great task. So I used a copilot function, which is in beta. And so how many times did you have to match a dirty list with the master list and try to figure out which of these 15 companies here are linked to these two different names. How many times did you do that over your career, Glenn?
Co-Host: Glenn Hopper(30:33):
Oh my gosh. I mean just part of you just knew that any day of the week that could happen.
Host: Paul Barnhurst (30:40):
So I did this for an example with a makeup company. And so they had 25 brands. I created a list of 125 with messed up names. So the brand name would have ink at the end and the next one would have ink with the period and incorporated in company and two spaces instead of one. And everything capitalised. And I just put a prompt in copilot that said, look at this list, look at the master list return the master list that matches 145. Took a second. All right, got every single one correct. There's one time I got a row off, I had to run it a couple times, but now that was early. It's still a beta function. It got it. All right. Sure, I could do that in the agent mode and have it built in my worksheet, but I can also use the copilot function, which is, that's a cool use case.
(31:26):
So cleaning data for another example I did is I took, and I think this gets to the ROI, so I took some bad data and I could even pull it up, but we'll just kind of talk to it here. And so what I had is I had transactions and sometimes they were text, sometimes they were numbers. I had names where sometimes they're all capitalised, sometimes they had extra spaces. I had numbers. Well, sometimes they had a comma in them, sometimes they didn't. And I used a copilot function, cleaned it up in like a second. I also used agent and it created a new table. Funny enough, I used Microsoft's clean data feature that uses copilot and it did the worst job of all of 'em because it only has three options of what it actually cleans. And then I did it manually so people could see. And it was really interesting. Agent copilot function were definitely the best too in this exercise. So cleaning data is a great use case. Formulas, you're reviewing your models, helping you with logic. You can even help it build scaffolding, headcount, schedule deferred revenue graphs. I haven't been, I'll give an example, waterfall chart. I have still not been able to get a single tool and I got to prompt it a little bit more to actually set beginning and ending bar as a total in Excel.
Co-Host: Glenn Hopper(32:43):
Waterfall charts have become the bane of my existence. I don't know what it is. Ai, not ai. They're just such a pain to do and it's always that problem. It's like, no, these all add up to the same total and it's an AI just cannot get that. But endless who can
Host: Paul Barnhurst (32:59):
Get it either. I'm glad I'm not alone. So graphs in general, what I've found is if you want it to build something and you know what you're doing, you can clean it up fine. It can give you a base start, but anything fancy, it's going to struggle. So great for giving you ideas for charts, great for reviewing them. That's not neat. You can do that with an LLM as well. But it's kind of nice. You have your data right there. You're write in Excel, you're not upload it to the LLM, wait for it to come back. So I think one ROI is just the latency, the time difference between going out to the LM every time and not going out to the LLM. The other benefits is review my formulas, help me write formulas, clean my data. You could help with some basic variance commentary. I've had to do PVM several times. It's pretty good at building a deferred revenue schedule so you can definitely use it. Those are the things I'm seeing. And I would recommend every finance department, every FP&A department should have some kind of agent in either Google Sheets or Excel that they're playing with and that they're learning and they're figuring out how to be more efficient.
Co-Host: Glenn Hopper(34:12):
I agree with that. And I'm having sort of this same existential crisis that so many professions are right now. These agents or the models themselves, whether it's Claude Cowork or improvements in other models and open source models that are out there. The issue to me is even if they're not right yet, even if it's extra work for now, I talk to FB and a teams and companies a lot and so many of them are like, eh, I do what I do in Excel and I don't really care about ai, I don't need it. And then there's some people on the team, whether they're admitting it or not, are using the heck out of it, but they might be doing everything perfect and it might be great, or they might be doing really dumb stuff and passing off AI work as their own that whatever, that's between them and their manager.
(35:04):
But if AI work is wrong and they can't explain it, and you've got just this black box where it's all going. But you asked about building ag agentic workflows now, and there's kind of a crisis around workflows right now because the innate end community that I've been pretty active in for the past couple of years, everybody now is thinking, well co-work can do half of what we were doing in N eight N and it's only going to get better. And so the idea that you have to go and N eight N is like make or Zapier, these other tools out there where it's a low-code, no-code, drag and drop kind of thing. They've even got an interface now where you can talk to an agent and tell it what you want to build and it'll get you 80% of the way there. But if you're not familiar with how to prompt them, what it is you're doing, like building agents is more about process thinking and engineering than it is about the actual finance knowledge. It's like, yes, I understand these are the steps. Here's our month in close schedule. This is what we do. Here's
Host: Paul Barnhurst (36:12):
Our process, here's our checklist, not process occupant, but here's our checklist. Let's just say.
Co-Host: Glenn Hopper(36:17):
And I would think any team I ever had, they had to build SOPs for how to do it. You hit by a bus or win the lottery or called up in the rapture or whatever happens to you, somebody needs to jump. Yeah,
Host: Paul Barnhurst (36:32):
Bus I'm not a fan of.
Co-Host: Glenn Hopper(36:34):
But the thing with that process thinking and there's tools out there now. I mean think of Scribe or these other tools that make creating SOP videos and make things easier, but these standard operating procedure documents are the playbook for getting work done. I think about all these automation tools. To my mind it's just building a flow chart in Vizio that actually does what you're saying for it to do. I'm hard pressed as long as there's an API and even if there's not a supplied API, well if there's not a supplied MCP or piece of software, as long as there's an API connection, I can move the data across the cloud from one system to another. I think I'd be hard pressed to not be able to automate any digital task that's out there. I mean that's kind of where we've gotten. And interestingly, I do a LinkedIn learning courses and the last course I did for them, they told me, Hey, your high level overviews are great, but we're leaning more into build courses because people want to build stuff.
(37:39):
Well, I did a how to build variance analysis sort of budget to actual workflow that had three different agents in it. One of them looked at trends, one of 'em did horizontal analysis. One of 'em looked at performance to budget. And it's a pretty cool tool that you can learn how to build in an hour. And that course is performing far worse than any of my other courses. And I wonder if it's not because yeah, you could learn how to use a new piece of software even though it's low code, no code, there's a learning curve. Or I could just wait for AI to catch up and be able to explain this process and have it automated. For me. I've been playing around because with my executive assistant, I built SOPs for everything for what they would do. And I'm wondering now can Claude do a lot of this?
(38:24):
And it can't quite yet. But I guess to your original question on what the mindset and how you need to be thinking to build agen workflows. The first is data governance, data dictionary. If you don't have the right data and the right access to it, understand your source of truth, then it's worthless. But after that, it's what you were just talking about of playing around with the tools, seeing what you can do with the tools in Excel today. Even if you're not doing anything with real company data, understand how these models work and interact and then think about that process flow and think about everything in terms of aviso diagram. What's the workflow we follow? Where do things drop out? Where are the exceptions? How do I need to solve this? And that's going to be in the not too far distant future if it's not already here.
(39:15):
In a lot of ways that's going to be all you need to do to build an agent to automate your tasks. I talk to companies all the time about these big massive projects that we're doing to tie together systems and transform data and do amazing things. Meanwhile, they're employees who have chat GPT or Claude or Gemini or Copilot or whatever have basically figured out this is like RPA for me, this is my own RPAI can automate my tasks. The smarter ones are. So really it's more rather than here's my playbook for how to build an agent, it's really about experimentation and shifting the way you think and the way you look at the tasks you do every day.
Host: Paul Barnhurst (39:53):
Interesting. So let's say they want to build their first agentic workflow. Is there a tool you would recommend? I mean, should they try something like an N eight N? Should they try it with chat GPT? Should they try these co-pilot that they're calling agents?
Co-Host: Glenn Hopper(40:11):
I was first really big, so Zapier was great for connecting early on, just connecting via API and doing things between Slack and your email and the connections there. And then I shifted to make, I can't remember why. Maybe I just wanted to see something different. When I found N eight N, it to me somehow was the easiest and more intuitive. And I hate to, I know I've thrown out a bunch of software. I'm not necessarily, I guess I did give a pretty big endorsement of Claude, but across the board, I like N Innate N. But I think Make and Zapier, and there's a couple other tools out there are pretty close to it. It's whatever you're comfortable with, but think of it as, don't even worry how it's working. If you can make a flow chart or if you ever built anything in Vizio, just take some process you do every day and map it out and say, here's a decision point.
(41:06):
What happens at this decision point? And here's I go to this system and then I do that. And once you map that out, well then you dump it into your favourite LLM and you say you don't even have to draw the figures. You can be like, these are the steps I take. Outline this workflow for me and then it will put the order around it and everything and then you could have it help build. But then once you have that, you could either go into one of these tools and build it yourself or now increasingly more and more you could just put it into your favourite LLM and say, okay, design this workflow for me and right now chat, GPT and Claude will all try to write the JSON that you can copy and paste into N eight N or one of the others. But they're not very good at it.
(41:46):
So I would say, I think all these tools now have their own agents that build for 'em. I would say get your prompt refined in whatever your favourite LLM is and then dump it into one of those automated tools. It's going to have hiccups and glitches, I would bet. And that's where the domain expertise and human insight and ability would happen. But it's really just about thinking about what that process is in trying low-code, no-code tools. And then OpenAI actually did roll out something. It sort of fizzled, not a lot of fanfare, but I think it was back in October, but I can't even remember what they were calling it. It was workflows and they did the release on it and I never heard anything about it. So pick your favourite tool, whatever you're most comfortable in, use the agent to help you use the LLM to kind of guide you around it. But think about data, what you're doing with it and that process thinking what is happening in each steps and that's going to get you significantly there towards building the agent. Maybe that sounds like I've oversimplified, but that's where we are with the technology right now. With that much and the help of an LLM, you can get
Host: Paul Barnhurst (42:54):
There. Yeah, that's what I'm hearing. And that's my next experiment. I'm spending so much time on playing with the Excel agents. I haven't really done that, but I keep thinking I can automate this or that, but I have an assistant who does a lot of things and I'm like, all it's working. I could focus elsewhere. But starting to just see that, to me, that's kind of the next level of skill that finance people need. If you're not doing it, you need to be doing it great. You've learned how to prompt, you figured out how to get some benefit from the A LLM. Now figure out how to use the Excel agents or whatever you want to call it, but also be willing to experiment with an agentic workflow, figure out how to automate something and it's going to take some time. First time it's going to take longer than the process takes to do. That's normal. That's no different than when you automate it in Excel or you automated anywhere. The first time may take as many as the first year, but if it's taking it five hours a time, you're going to save it over the long run. If it's a one-time thing that's different. But if it's a monthly or a weekly, the savings will come and you'll get better and better. Fair statement
Co-Host: Glenn Hopper(44:01):
A hundred percent. You don't want to be, when I had my first person that ever reported to me very early in my finance career, and he was over procurement and was an older fella, and he had, this was back in the days of CRT monitors, big monitors on the desk and his was never turned on. He did all of his tracking. I'm old, I'm not that old. This would've been early two thousands. Dude never turned on his computer. He had a paper ledger where he tracked everything. So he wasn't really moving up in his career. But by not embracing AI right now, I'm not saying, and I know I can sound like an evangelist on it, I'm not saying throw caution to the wind. And you see so many people out there posting on social media about AI does all these miracle things. And I think yes, it can do great things, but at the same time, I'm not giving it the keys to the kingdom and taking myself out of the loop and letting it run the business. However, erring on the other side is maybe even, well, I don't know which is more detrimental. That's a whole other debate. It's
Host: Paul Barnhurst (45:05):
Dangerous in and of itself.
Co-Host: Glenn Hopper(45:07):
So I would say you've got to experiment with it. And you can't be the one who's left behind by this and you a hundred percent. We had to change our skillset when spreadsheets came out, when desktop computing came out, when SaaS, I mean that's just the nature of business. You have to evolve with it. So you get very frustrated when I get the Luddite responses around ai, which you see a lot in finance, accounting. And I'm going to go ahead and make another dig at the accounting folk. You see it more on the accounting side than you do on the finance side. So as we wrap up, you've
Host: Paul Barnhurst (45:37):
Made a big bet on Claude. You've said right now, that's the best tool for me. This is not a promotion saying everybody go run out and use Claude for everything, but it's got a lot better on the commercial side. Your bet is GPT going forward is going to be more of a consumer tool so that we've covered that too. The advice on Excel agents is be using them experiment. They're great for formulas. They can help with auditing, they can build things they can help with what charge to use. They still have a lot of errors. They're definitely not perfect. It's very much human led. AI assisted is the term I like to use, and they're a magnifier. Now when it comes to workflows, and we're calling it agentic workflows versus agents, multilevel reasoning, but really writing tasks, those operating procedures, get that operating procedure. Make sure you understand your data.
(46:32):
Get into the LLM and have a help you get those steps, all codified, the image, whatever. Then take that prompt and go to one of these tools and eight in make zir whatever they may be, one of these no-code automation tools. And try automating that first process. Take something that's simple, start there, start doing a couple, and you're going to have a new level of where you are at in your progression. That's kind of the next level. I need to do more of that. Right now I'm testing these Excel agents and I'm like, I need to do some ENT workflow stuff. I really haven't done it, I'll be honest. And I am talking on a podcast all the time about ai, so I need to do better myself. Glenn's done many of 'em. I'll probably pick 'em up and go, Glenn, I screwed up. Help me. So that's kind of our summary today. Last thing we'd like to say is if you've listened this long, email us, we'd love to hear from you. You can reach out to us. I'm p barnhurst at the FP&A guy.com and would love to hear your questions so we can have another mailbag episode. So there's my thoughts. Let you give last word, Glenn.
Co-Host: Glenn Hopper(47:35):
Yeah, I think you very succinctly put it and not really a lot to add to that. I will say if you're listening to this podcast, chances are that you're already pretty deep in the AI world, but don't stop experimenting and it's moving faster than any technology I've ever seen. So thank you for listening and hopefully we've given you some good ideas and tips for going forward. And like Paul said, we'd love to hear from you and hear what's working with AI with you and what's not. You're
Host: Paul Barnhurst (48:02):
Right. Well, thanks again for joining us and have a great day. Thanks for listening to the Future Finance Show. And thanks to our sponsor, QFlow.ai. If you enjoyed this episode, please leave a rating and review on your podcast platform of choice, and may your robot overlords be with you.
