# 167 I built an AI agent finance department (full build)

Show Notes

I brought on Mike Dion, a senior corporate finance professional who has helped automate more than 100,000 hours of work out of finance processes, to walk through how to build a fully functional AI agent finance department inside N8n — from scratch, with no coding background required. We build a CFO agent named Charles, give him a team of specialist sub-agents (FP&A, Accounting, and Treasury), and watch the delegation logic in action as Charles routes questions to the right specialist instead of answering everything himself. Mike also breaks down how to use ChatGPT to write your own system prompts, why you should give the CFO a reasoning model while using cheaper models for the sub-agents, and how to publish the finished chat so your whole team can access it. By the end of this episode, you'll have everything you need to replicate this AI finance department in under 45 minutes.

Timestamps

00:00 – Intro

00:03 – What we're building: AI CFO inside N8n

00:23 – Mike's background and 100,000 hours of automation

00:52 – Why N8n over Make, Zapier, or Power Automate

02:00 – Setting up the chat trigger and naming the CFO

05:00 – Using ChatGPT to write the CFO system prompt

08:52 – Choosing the right AI model and saving on token costs

13:27 – Building the FP&A sub-agent (and what FP&A actually does)

16:36 – Adding the Accounting agent with code interpreter

19:25 – Building the Treasury agent

33:21 – Successful routing to FP&A and Treasury agents

38:19 – Publishing the chat and embedding it in Slack, Teams, or a website

39:22 – Mike's philosophy: train your team to build, don't just build for them

41:14 – Where to find Mike and his free Finance Automation Insider newsletter


Key Points

Using ChatGPT to write your own N8n system prompts is one of the fastest ways to get started — nothing knows ChatGPT better than ChatGPT itself, and what would have taken six hours of writing two years ago now takes minutes.

The delegation logic is non-negotiable. If the CFO answers questions directly instead of routing them to a specialist, you lose access to any tools or context you've connected to those specialist agents — and you pay more for it.

N8n can run completely free on a $4–5/month virtual private server, making this entire AI finance department buildable for nearly zero cost. You don't need a paid automation platform subscription.

Setting the context window to 10 (five back-and-forth interactions) is a practical default — enough for most finance questions without ballooning your API costs on every run.

You can publish the finished CFO chat and embed it directly in Slack, Microsoft Teams, or a company website. Your team sees a clean chat interface — all the N8n complexity stays invisible in the background.

Links:

F9 Finance YouTube channel — Mike's free weekly channel covering finance automation tools and builds: https://www.youtube.com/@F9Finance

F9 Finance website — corporate automation training and the free Finance Automation Insider newsletter (includes 15 five-minute finance automations you can build with tools you already have): https://f9finance.com

N8n — the free, self-hostable workflow automation tool used to build the AI CFO in this episode: https://n8n.io

Join the Build With AI community — weekly AI implementations, live coaching, and no-fluff templates built for non-technical entrepreneurs: https://www.skool.com/buildwithai/about

If this episode was valuable to you, it would mean a lot if you left a rating and review on Apple Podcasts or Spotify. It helps more entrepreneurs find the show.

FIND ME ON SOCIAL
X/Twitter: https://x.com/coreyganim
Instagram: https://www.instagram.com/coreyganim/
LinkedIn: https://www.linkedin.com/in/coreyganim/
YouTube: https://www.youtube.com/@coreyganim


FIND MIKE ON SOCIAL
YouTube: https://www.youtube.com/@F9Finance
Website: https://f9finance.com

What is Build With AI?

Most AI podcasts talk about what's possible. Build With AI shows you how it's done, live. Each episode, host Corey Ganim brings on entrepreneurs and operators who share their screen and build real AI automations, workflows, and tool setups right in front of you. No boring slides. Nothing that hasn't been battle-tested. You'll watch actual implementations get built from scratch so you can follow along and do the same in your business. If you're a non-technical entrepreneur who wants to put AI to work without becoming a developer, hit play and build along with us.

Corey Ganim: Mike, what are we going to learn today?

Mike Dion: we are gonna walk through how to build an entire finance department and a CFO inside N8n, the automation tool.

Corey Ganim: That is so cool. So we're going to basically be building an AI agent CFO inside an eight end. I'm excited to jump in and see this thing in action. So before you share your screen, what is your background? Like, why should people listen to you?

Mike Dion: I am a senior corporate finance professional. I've been working in corporate finance for over the last decade. And across that time, I've helped people automate more than 100,000 hours of work out of their processes. And today my mission is to help people double their impact in their companies while working 40 hours or more or less a month.

Corey Ganim: I love it. Well, that's so cool. So let's go ahead and dive right into your screen and we're going to break down this AICFO step by step and see what's under the hood.

Mike Dion: Sounds like a plan. So as I pop this open, so as I mentioned, we'll be working inside of N8n. For anyone not familiar with it, it's a very similar tool to things like Make, Zapier, Power Automate, essentially a workflow tool. The great thing about N8n, first of all, if you're willing to do a little bit of setup, you can actually run it completely free, whereas those other tools will drain your pocketbook real fast. ⁓ I run this ⁓ on a virtual private server on my website. It's all really easy. That's like, four or five bucks a month and then ended in itself is free. And then the other thing is just a lot more customizable and dynamic Zapier and make or a little faster to set up, but you can basically build anything with any then. And if you want to get really fancy and start coding, you can build even more with any then. But I'm going to show you in this video, you don't actually have to write any of the code yourself. We're going to build this entire CFO with AI. So it's going to do literally everything for us as we go. And that's really the sweet part about this is regular everyday people who are only used to working in tools like Excel can now build things that previously only tech developers could do and it's really exciting time.

Corey Ganim: Yes. I'm so excited to get into this because it, and it ends powerful too, man. Like in a den super powerful. And I think a lot of people are intimidated by it. So I'm excited to see how yours is set up because I mean, correct me if I'm wrong, but you don't have like a development background, right? Like you're not a developer or you're not a coder necessarily. Like you're a finance guy.

Mike Dion: So that is N8N. Not remotely, the only coding I ever did professionally was VBA and Excel, which is barely coding. Maybe a little bit of a little bit of HTML tossed in there from just learning how to do some websites, you know, in high school and early on. All of my coding expertise has been self-taught working in partnership with AI. And I would say in partnership, because I think it's really important. You can't get in the mentality that you just have AI do everything. You need to learn as you go. You need to understand what you're doing. You need to understand the framework of how code works, but you can lean really heavily on it to assist you. So partnership is really what it's all about. So let's get rolling with this. let's go ahead. We'll start where we should always start. We'll just give this a nice title like to stay organized as we work. So this will be our AI CFO. For anyone not familiar with workflow tools, everything starts with a trigger. So there's a lot of different choices of trigger in N8n. There's some really fancy ones that can involve web hooks and things like that. We're going to have this start very simply though, and we're going to build this like it's a chat. So our users at the end of the day will interact with this CFO. in chat form, is how you would interact with the CFO. It's it's going to be like just slacking or teamsing with a CFO. We're going to set this up later, but I'm going to toggle now to make the chat publicly available. And what that's going to let us do is two things. I'm going to show this to you essentially hosted in N8n, but you can also embed it on websites, or you can embed it in company tools like a ⁓ Slack or a OneDrive or things like that. It basically turns what you're building into an application that other people can use. It's great to build automations for yourself. It's really powerful to build automations that an entire team of people can use. And then we'll just customize the message users will see. So say hi there. I always call CFOs Charles. It just it's a CFO name, you know. So my name is Charles.

Corey Ganim: So I was about to say, was like, that just sounds like a finance name. Maybe it's cause of like Charles Schwab.

Mike Dion: It does. It does. I am your. Chief financial officer. All right, so that's what it'll say. So when the users come in and we'll set this up at the end, once we have it ready, they'll come in, they'll chat with this, just like it's your CFO named Charles, and he's gonna answer all your questions and solve all your problems. So we'll kick that off so it's ready to go. And that's the start. So that triggers it. You can set off your automations by time. You can set them off to trigger when another sequence runs, lots of options, but we're gonna keep this simple and do it as a chat. And then we're gonna go straight in to the core power of this tool. Now I mentioned we're gonna build a CFO and their entire finance team. So we're gonna start with a really powerful AI agent that is the CFO. That CFO is then gonna hand off different tasks. If it's an accounting question, they're gonna go to their head of accounting. If it's a treasury question, they're gonna go to their head of treasury. They're gonna be able to hand it off to all those different people. But our first agent is gonna be the really powerful CFO. So we're gonna go to our AI tools. We're gonna stick an AI agent in. And then we're gonna start building this one. Now what's really critical with this is getting a good system prompt. For people who are not familiar with system prompts, this is as you're working with AI tools that you wanna be reusable. It teaches the AI how to behave. It's essentially its brain and how it's gonna approach tasks and think about it. Now, writing really good system prompts can take a while, but we're gonna actually have ChatGPT do it for us. The reason I'm using ChatGPT here, first of all, it's just kind of one of the tools I default to, but we're gonna be connecting all of this to OpenAI models on the back end. That's what's gonna drive this. So we're gonna have ChatGPT write the system prompt because nothing knows ChatGPT better than ChatGPT itself. So we're gonna ask it to help us build it.

Corey Ganim: Right.

Mike Dion: I'm going to pop chat GPT over here. I am in pro. You don't need to be in pro to do this. Everything I'm going to show users can do with the free version. I just have pro and I'm using it for that. The only real difference is if you mess up and want to try again, you get less prompts with the free version a day than you do with the pro version. Aside from that, everything you're about to see can be done free. It can also be done in Claude, Copilot, Gemini. Really everything I'm doing today can be done with any of the tools and you can just swap them out one for one with chat GPT. So we will paste in a prompt. We're gonna have it act as an AI agent expert. We want to help building an AI CFO on N8n. It's gonna be using an AI agent node with thinking and reasoning capabilities. It's gonna call on other AI agents. This is gonna be FPNA, accounting and treasury. It's gonna process requests. And we're gonna ask it to help us write the system prompt for this AI tool. I will say ChatGBT has been overdoing it lately with the intros a little bit.

Corey Ganim: And I'm so excited to. Yeah. ⁓ yeah. I feel like it can get a little, it can get to be a little verbose, but I was just saying like, I'm so excited to kind of work through this whole process, but it's so, so timely for me because I'm, you know, I've been messing around with open clock quite a bit and I've been thinking about like, how do I have my own financial, whether it's an open call agent or in this case, I'm probably just going to go rewatch this video and just duplicate your workflow and N8N to have my own, ⁓ agent CFO. So I'm really excited to get in the weeds here. I'm already like,

Mike Dion: But sorry you were saying.

Corey Ganim: My head is already spinning in terms of ideas here.

Mike Dion: Yeah, and I'm going to be showing a pretty straightforward setup because I like it. You know, it's I like to be a little bit more approachable. But as you play with it and want to start adding features, I mean, you can really put some put some fancy stuff into these. You can build all kinds you can start building in, you know, it can start querying, you know, rates from foreign exchange websites. It can start querying interest rates. I you can connect it to all kinds of different things and really, really neat stuff. So here is our system prompt. We'll go ahead and copy this over. We're gonna put this right down here as our system message. So this is what's gonna give the tool its brain and then the prompt. So we don't need to prompt it. The prompt is gonna come from the chat and that's why you see this chat input here. Now you'll notice this has the dollar sign JSON. If you're not familiar with JSON, it's a coding language. It's a simple coding language that really just is a way to structure data that AI tools and other computer systems can handle it easier. You don't really need to know that because AI can help you. If you have any questions, if you ever get stuck building automations, just ask AI to help you, ask it to explain it to you. It'll walk you through really easy. But just know that JSON is just a way to just structure data so the AI tool can read it. And a lot of it is going to write itself as we work through NADAM. So you'll see that. Nothing to get scared about. Again, if you ever have a question, copy, paste it to gpt or Claude, it will answer the question for you. So this is the core of the agent and it's built. Now what we need to do is we need to enable it to work. So right now it's literally just a system prompt and the ability to take in chats. We need to give it the actual AI model. We need to give it some memory and we need to give it some reasoning power so it can work through different tasks for us. So let's start with the chat model. So we're gonna come down here to chat model. All of the options you could possibly want are here. We're going to just work with open AI since that's what most people are familiar with. And this is a great place to talk about how to save money with AI tools. So right when you're working in chat GPT, you're not really burdened by the amount of data you're sending back and forth. But when you start connecting to AI, you're paying for every single token and a token is just a bit a little piece of what you're sending back and forth. It's almost like a syllable. And you want to think carefully about that because if you're going to run these a lot, it can get pricey over time if you don't structure it right. One of the benefits of this system is we can have the AI CFO use a more supercharged reasoning model and then have most of the work done by cheaper models in the in the agents, right? We don't need to give them all the same strength, so we're going to give a reasoning model to the core CFO, but we're going to give kind of just the basic. regular old chat GPT mini models to the tools and that's going to give us a lot better capabilities to manage our spend and put the abilities in the right places, right? Don't don't supercharge it because you're paying for it whether you need it or not. So we're gonna go ahead and scroll down to 4.0. So there's GPT 4.0. That's gonna be a little bit more of a powerful reasoning model. And since this is gonna be a CFO, so I'm not gonna do web search or file right now, but I'm gonna give it code interpreter because code interpreter will let us work with numbers and do a lot more reasoning in that area. So we'll turn that on as well. All right, and that is our model. So this is what it's gonna call on when we ask it any questions. and that part is done. And you'll see as we add, it's basically building out this little workflow that we can follow along with. You just follow the arrows, you follow your workflow. If it's a dotted line, it's not in the core workflow, it can just be called on by the AI agent. The AI agent is the real trigger. Memory, you can get fancy. You can add all kinds of external structured memories. We're just going to do a simple memory right now. It's going to store it inside of my N8n instance. So we'll pop that in context window. This is how far back in the chat you can look. So we're going to give this. We're going to let it look back 10 chats. Don't want to go too many or it'll start start costing you a lot because it's going to send this back and forth to open AI every time, but we want to give it enough that you can have a little bit of a back and forth with the CFO. that takes care of the memory. And then our first tool we're gonna add is gonna be reasoning, which is gonna let this tool do a little bit more thinking than the average agent would. This is gonna be in AI tools, because I clicked on memory. AI agent tool. Let's go down to reasoning. Thinking, thinking tool, not reasoning. That's the problem. Thinking thinking think tool there we go

Corey Ganim: And just look how many tools it has access to. mean, there's, there's so many different capabilities you can give the agent. Right. So for folks who haven't worked inside in eight in before, if this is your first time kind of jumping into a. A node based workflow builder. I mean, the possibilities are endless with this type of stuff, which is what's so exciting about it. So just context for people ⁓ watching or listening to maybe haven't jumped into innate in before there's so many tools you can equip your agents and your workflows with.

Mike Dion: Yeah, it's true. mean, there's over 250 pre-built connectors. Plus, there's the ability to connect to anything that has an API. When you add it, when you go to add a node, you'll see all these different options. But within each of these, like if I go to action in an app, you're gonna get this list of more than 200 actions. I mean, there's a very good chance any tools you're using are in here or have the ability to build your own connection. I mean, it's just fantastic. So this guy right here, this is our CFO. This is our Charles right here. And now we need to give our CFO a team of agents. And for that, we're gonna give them in this case, a tool and that tool is gonna be another agent. If you wanna get really crazy kind of inception style, you can even give your agent tools, other agent tools, and you can just kind of keep going down that. You'll burn through a lot of tokens and spending, but you can just keep going and give every agent their own agent and. and go crazy, but we're just gonna do one layer here to keep it simple. So we'll go ahead, we're gonna add in another agent. Here's our AI agent tool. All right, and now this is gonna be where we put in, we'll start with our FPNA agent, then we'll do accounting, and then we'll do treasury. And just like we did for the AI CFO, we're gonna go back to chat GPT and ask us to help it with this. We'll keep this prompt pretty simple because we've already defined the task and we've already defined the role and we're still well within the context window. So we'll say continuing your role help me build the FPNA agent. I'll put a description and a system prompt so this is going to start helping us build out all the sub tools and then we'll use that same prompt again for Treasury and that same prompt again for. The FPNA agent is not a spreadsheet monkey. That's so funny.

Corey Ganim: So this ⁓ is a dumb question for someone who, like myself, has no finance background. Like, what does FP &A mean?

Mike Dion: That is a really good question. So FP &A is financial planning and analysis. So the best way to think about it is accounting looks backwards. Their job is to keep all the books straight to make sure that you're reporting correctly to the IRS, the SEC. FP &A's job is to look forward. So their job is to plan where the business is going, to analyze and provide insights about where the business is today, using accounting's data of kind of what's happened, but that's the differences forward-looking versus backwards-looking. There are two separate teams under the CFO, one of which is just focused on getting the actuals and the books correct, one of which is focused on kind of guiding the organization. All right, so here's our agent description. Now this is probably a little bit more than we needed, because the description is, it does inform the agent a little bit, but it's more so for the users. So we'll just take in the top sentence here. All so that'll be our agent. And then just like before, we'll add a system message. And here is our system prompt for the FPNA agent. All right, and then just like the core CFO agent, we need to give this one capabilities as well. So let's try to continue keeping us organized. Slide this up here. So we need to give it again a chat model, a memory, and then any tools we want, which will in this case be thinking. So our chat model, we're gonna still use OpenAI, but we're gonna use a lighter model that's not gonna take up as many resources. Cut our spending down a little bit. We'll just go pop in GPT-5-2. This is the current working model in ChatGPT. That's a pretty good guide. If you're trying to do just regular basic work, the core model at the time will be one of the more newer, more advanced, but won't cost you too much in spending because it's not one of the really supercharged models. So just a general rule of thumb, if you want to keep track of all, mean, there's 90 or so different models in here. If you don't want to try to figure out what they all mean, just look what's in ChatGPT at the time. That's probably a pretty good one to be using. Again, we're going to give this the ability to interpret code. Coding in this case gives the ability because a lot of the interpretation of financials is going to be done with Python or R and this gives it that ability. So it really works a lot better with that information if you give it the coding ability.

Corey Ganim: So that's why I was going to ask. I was like, so if we're dealing with finances, why exactly do we need to give it a code interpreter skill? But that makes sense when you explain it that way.

Mike Dion: Yeah, and you know, it's always important to remember the tools are LLMs, they're large language models. These AI tools evolved to work with language, not with numbers. They can work with numbers now. So if you went back to like GPT-3, early fours, they were really bad at numbers. They hallucinated. They couldn't, you know, if you ask them one plus one, they tell you it's four. ⁓ That's because they hadn't developed kind of these reasoning capabilities or the ability to work with things like Python as well. All of that's gotten better because if you're in chat GPT and using it, it just turns these tools on naturally. You don't need to tell it to do it because we're designing it ourselves. We need to tell it to do it. So that's why like you'll see chat GPT do this automatically and it's totally fine. We're kind of developing this so we have to help it a little bit along the way. And again, anytime you know we might even get an error or something we work through this or not like our output. We're just going to tell chat GPT where we're having an issue and it will help us rewrite our system prompts. You don't have to get it right first time, right? That's the biggest thing is no one's seeing this until you let them see it. You don't have to be perfect. And then we'll just go ahead and add our think tool. There's that, and then we can give this one a name so we remember what we're. Now we're going to that exact same thing. We're just going to do it more times for our other tools. So I'm back to chat GPT. Take our prompt, turn this into an accounting prompt. and we'll fire that away. So what do you think? How are we looking so far?

Corey Ganim: I think we're looking great. This is awesome. I mean, again, it's just what I like about what you're showing us as you're breaking it down step by step. Like anybody watching this or listening really could go and replicate this. And we're only 20 minutes in and you know, I'm sure within the next 15 minutes or so, they could have a full functioning system exactly like what you're showing us. And you know, none of it requires any sophisticated knowledge or paid tools. mean, you have Chad GPT Pro, but you even said you don't need it. Like you could get this rolling for a couple bucks, right?

Mike Dion: Absolutely. Yeah, I think I always want to make sure like I too too often this tech work is made to be a little bit more intimidating than it really needs to be. And two years ago, even it was I could not have been building an automation like this two years ago without a lot of time and effort I didn't have. But the ability of AI tools to teach us anything and guide us through. The biggest thing I'd say if you as you're learning this is don't try to build something crazy the first time.

Corey Ganim: Right.

Mike Dion: It's still important. You can't outsource your learning or your understanding to AI. You can outsource the technicalities to AI. You want to understand what a workflow is. You want to understand your process. You want to understand what AI tools can do. But you don't have to understand how to write the entire system prompt. You don't have to understand how to generate every line of the Python. You want to understand the structure and the fundamentals. And you can even ask a how to teach you that. So don't skip that step, right? It's still not the time to do that, but it is the time to let it. Batch I mean, I think how long it would have taken even two years ago to write all these system prompts out. This could have been six hours of work to write these.

Corey Ganim: ⁓ yeah, it would have taken forever. Yeah.

Mike Dion: So there's that again, we're going to give it our move this up here so we can keep organized. Let's go ahead and give it our chat model. Give it 5, 2 again. Give it a code interpreter. memory. ⁓

Corey Ganim: And so is the default context window length, do you always set it to 10, which basically just means like, Hey, it's going to reference the 10 most recent chat messages from the user, ⁓ before giving its response. Is that what that means?

Mike Dion: Yeah, so what I like, it's five back and forths. So it includes the AI's response. So the reason I do 10 is it gives me five interactions. That covers most questions you'd be asking. ⁓ Five literally means if you go back and forth twice, you're out of context. And that's just.

Corey Ganim: Okay. Got it. Right. Okay, that makes sense. Got it. Yeah. So, so theoretically it should always be an even number.

Mike Dion: Yeah. And full production wise, you probably wanna even go higher than 10, but for purposes of this, I mean, we're just gonna ask it one or two questions as an example and you'll get it. It's really depending on ⁓ if you're paying for it, if your employer's paying for it out of kind of a pool of tokens, depends on just kind of the capabilities you wanna give it.

Corey Ganim: Right.

Mike Dion: There is oh and I'm remembering. Let me go ahead and run this. I need to get a little bit of. So what I'm gonna do right now that you're gonna see is I'm going to let the AI agent run because I need a piece of that JSON code that's passing through. I need to be able to drag and drop in these tools and that's why we see the little thing there. So a nice easy way to just get it to run is say, what can you help me with? That's gonna give it just enough to get going. It's gonna give us this ID we need inside the code. And it's this session ID is what we needed, so I just needed something to go in so it could run for me. So let me go back and fix this error we're seeing. So here we want to just pass down the original chat, right? So we're to take that. And this is the great thing is again, you don't need to know the coding if it's already over here, you can just kind of drag and drop it so it's a little bit easier than it even looks on first glance. We'll go back down to do the same thing here. Add input, pop that over. did not successfully save the name, so we'll rename this. All right, and then the last thing we need to do is just build our treasury agent and then we'll be good to do a couple of demos. So now we'll add one more tool and we're just gonna do the exact same process again. Now in playing with this, say so I've actually built this with as many as eight AI agents. There is absolutely no reason that everybody needs to sit here and watch me build eight. Three is more than enough, but there's no limit. You can keep going as many as you want. You can have as many individual tools as you want, ⁓ but it is reiterative. mean, it's just, it's just you redoing the same thing multiple times in a row. But once you get comfortable, you can, know, and if you hit any limits, that's the other thing you don't have to even try to guess where you need to expand this out to.

Corey Ganim: you

Mike Dion: You can get the basics in, you can get kind of, you know, to take from the tech world, minimum viable product, get it out there, get it working. The fastest way to learn where you need to build it out more is by getting it out there. So don't worry about trying to get every single possible scenario or situation covered, get the basics, get the most common. And then as you find yourself not getting what you need, you can adjust as you go.

Corey Ganim: Right, and that's really good advice. I would agree with that.

Mike Dion: Here's this, I will drag over our message, put on our system message. ⁓ we're still running. I'll give it a second. Got too ahead of myself. There we go. All right, we'll put our system message in. this are. Treasury agent. All right, so there's that. I'll pull it down here. Zoom out so we can keep seeing everything. Put it over here just a little closer. And then we'll do our chat model. Exactly the same as before, we'll pull in OpenAI. We'll pull in GPT-5-2. and we'll add our code interpreter. into our memory. Context, window 10. and give it our thinking tool. All right, and there we go. So this is an AI agent. We've done it. Just a couple of minutes, look at that. So now we're gonna test it. You never quite know what you're gonna get the first time you run it because until the system prompts talk to each other, sometimes it can go a little crazy. I also always love to point out that there's this rule when you're doing live demos that one time out of 10, AI just decides to do something insane and that one time out of 10 is usually when you're screen sharing. So always exciting to see.

Corey Ganim: That's awesome. Yeah. All right, it's Murphy's Law in action.

Mike Dion: Exactly always exciting to see what you get back, but that's why you know we have the AI tools here because if something goes crazy we can just give that feedback to chat. You PT and it will write a new one so let's start with a simple question. Now I'm going to put in a question that should be targeted to FP and A. So what is the best way to run a rolling forecast? That's a process FP and A is doing. So as we test this, what we're looking for is we're looking for our AI agent, our CFO to run. and then to choose just the FPNA agent tool to answer the question. So the process will be a fail if it doesn't pass off because we want it to be asking its specialist questions. The process will also be a fail if it picks the wrong tool. So a successful run of this would be the AI agent passing it to FPNA and coming back with a usable readable answer. If any of those things aren't true, we'll wanna refine our system prompt. All right, so we'll send that. So you see right now we're processing. So the CFO is working. When you see that little orange swirl, we're working. It's using the chat model. It's engaged its memory. And then we'll see what happens from here. The suspense is killing me.

Corey Ganim: I know I'm like, I'm really excited to see what it does.

Mike Dion: Okay, so we have an unsuccessful run, which is not too uncommon. The AI agent ran, the AI agent came back with a correct answer. And you might be saying, well, why is that a failure? It's because we don't want the CFO to be answering the questions. So it worked, it did give us an answer. So we know that at least it will come back with something, but we want the work done by the FPNA agent tool. So what we wanna go back to chat to BT and say is basically, you must, have the CFO either pass off to a tool or say that I don't have a specialist to answer the question because we wanna make sure it's being done by the lighter specialized model.

Corey Ganim: So basically giving it almost like delegation logic.

Mike Dion: But again, it did back with a cr- Exactly, It did this good. came back with a good, clean, correct answer, well-formed and easy to read. It just didn't behave the way we want this entire agent to work. Come back here. I just asked the model to do its first run. Oh, I have to use the available microphone. Let's hit that. I just asked the model to do a first run. I asked it a question about rolling forecasts and how to work with a rolling forecast. It did not pass this tool off to the FPNA agent. The AICFO ran it by itself. I need you to build in logic to the AICFO's system prompt such that it always passes off to one of its specialist agents or comes back and tells the user, it doesn't have a specialist that can answer that question. Please give me a new system prompt. And that is, as you're working with ChanTPT, one of the best ways to get information into the tools is to just talk to them. So if you're able to talk to them and you're in a place you can talk to them, I highly recommend prompting that way. They're, again, they're language models. They're designed to work with text and they're designed to work with written communication. And you can get your information in a lot faster than trying to sit there and type it out.

Corey Ganim: So yeah, so kind of like exactly what I was thinking it's now it's giving us a delegation and for system prompt. So previously it's, you know, if the regular AI CFO agent, obviously it's hooked up to the open AI model and it probably gets that system prompt and it's like, ⁓ I can just answer that myself. don't need to delegate that out to my, you know, FPNA sub agent, but what we're updating it with is we're updating the system prompt to say, you must. Right. You must either delegate or say that you can't answer, which is funny because that's kind of how you would want your human. You know, if there was a human in that position, you'd want them to probably operate the same way. It's like, Hey, either I have my specialist handle this task or it's something I just can't do. Like, of course you would want a human specialist to handle it. Same way you would want your AI specialist to handle it as well. I

Mike Dion: Exactly, yeah, there's two big reasons we want this to work this way. The first one is that it's cheaper, right? If you're running this over time, it'll be 30 to 40 % cheaper by having it pass off than by having the AICFL run it, because we're giving it all that reasoning power. ⁓ The second reason is just as you build this out, you can do things like you can connect your tools. I could connect my FPNA tools to my forecast, and I could ask it questions about my forecast.

Corey Ganim: Mm-hmm.

Mike Dion: And if I've connected the tool to FPNA, the CFO is not going to have direct access to that information because again, that's not how we built this tool. If it doesn't pass it off, I lose out on all that context that I've given it. So a little bit of his costs and a little bit of his just thinking ahead to make sure that the right AI agent is doing the right job with the right information. So now we'll give this another test run. So we'll try our rolling forecast question again. And now again, what we're looking for is to make sure that our AI agent passes off the task. So the CFO is running. and it passed off, look at that.

Corey Ganim: Nice, yep, it did exactly that.

Mike Dion: Yep, so now you'll see that the AI agent is still running. It's in basically paused mode. Only the AI chat is running for the FPNA agent. That's then going to stop at some point, turn green, pass it back to the AI CFO, and then we'll get our response. So that was a successful test as long as we get back a response that we like. There it goes, so now it's back to the AI agent. So now you'll see the chat model running again. It's basically formatting the response and it's adding like the executive polish. All right, so here's our response. So the domain classification is FPNA. It's routed this to an FPNA agent, and here is the specialist findings. So it gave us the findings, it gave us a summary, it gave us an executive overview what this means, risks, recommendations, and then saying for further assistance, you can tailor this to your specific needs and give me additional information. So that is a really successful run. Let's just give it another example for demonstration of going to a different agent.

Corey Ganim: Yeah, that's awesome.

Mike Dion: So now I'm gonna give it a question that should route to treasury. So we have excess cash on hand, that's treasury's function to manage cash. What are some good options for using funds until we have a need? And what we wanna see is exactly what happened in the last run, except using the treasury agent instead of using the FBNA agent. Now what's interesting here is we just had it basically lie to us because it says it routed it to the Treasury agent, but it did not. So let's let's let's yell at it a little bit. You did not pass it to the Treasury agent.

Corey Ganim: Yeah, I was like it didn't. Right.

Mike Dion: Please try again. So if we can poke at it and it behaves, we're good to go. There we go. So passed off. So it's a flag. If we see that behavior multiple times, we want to revise the system prompt. ⁓ What is likely happening is I probably should have just cleared the context. So bad on me for not doing that. I just got a little confused because there was no context clearing. And that's also, again, something that we could build in. If we saw it not behave once we poked it a little bit, then we would definitely want to re-write the system.

Corey Ganim: And so updating the system prompt in that case, like, let's say we, we kind of poked it and then it did the same thing again. Like it lied to us again and it didn't delegate properly. I'm assuming that that would involve like going into the system prompt and just, I guess, more strongly wording the delegation language of like, you absolutely must delegate, you know, at all costs, like just some, some just more extreme language to tell it like you have to delegate. to the correct agent or like, or else more or less, right?

Mike Dion: Yeah, I've I have found doing things like and it's weird it works. But if you do like capitals capitalized do not things like that the AI will respond really well to but we got a really good response. So it did route to Treasury. We have a specialist finding there's it's breaking it down even by like the length of cash kind of what we'd want to be doing where we can park it. All of this information looks great. And then it gives you kind of the executive bullets decision posture.

Corey Ganim: Right.

Mike Dion: and then I can provide additional details for a tailored recommendation. So again, a really successful run. And that's our AICFO.

Corey Ganim: That's awesome. I love it. And I mean, this is this system could apply to any small business. mean, any small business out there that is considering, you know, I don't think there's necessarily a, I mean, you might disagree with me. I think it makes sense to have like a human finance professional, maybe, you know, running this particular agent system here, but you certainly don't need like a finance team. If you're a small business, you could get away with One person in the business who wears a finance hat from time to time managing the system. I mean, that's going to get you to, would assume seven figures, multiple seven figures, maybe even eight figures per year in revenue that that system could support that. Like, has that kind of been your experience and installing this in your company and other people's companies and everything that the system can kind of provide?

Mike Dion: Yeah, I think at the entry level, you really see being able to replace a lot of the actual work. At management and more senior executive levels, what you see is it opens up kind of the brain space to do the things that really matter. What we want our teams doing, if you're in treasury, we want you spending time with banks, building relationships, getting better tools, finding better ways to park the cash, finding better ways to manage it, thinking through processes to help the company.

Corey Ganim: Right. Right.

Mike Dion: not having to answer all these random questions that come up. So what it's doing at that level, it can fully automate kind of the entry level work, but for management and executives, what it's allowing them to do is actually just spend time building and improving the business. And that's what we really want because there's not all these little tasks. And then just as an example, I pop this open. ⁓ I mentioned once this chat's built, you can publish it. You can now interact with it. This is just... had N8n create a small website for us. This is the chat. You can talk to Charles, our chief financial officer. You can embed this on a website. You can embed this in Teams and Slack, anywhere else that you're able to kind of embed chats. And it doesn't have to live in this technical N8n setup. You can format this to your own branding, to your own preferences, and make it fully available to your team and anyone who needs it. So it's not living in some system that everyone needs a login to. You can make it completely and easily available in this very... This is all your users see, it's just a regular AI chat. But you've done all this background work and customized it for your company and your purposes.

Corey Ganim: Now, is this something Mike that you build for clients or that you have built for clients? Like, is this one of your products like this, um, AI CFO that you do you sell this as a product?

Mike Dion: So I actually don't. So what I do, my method is after getting burned by a lot of really expensive management consulting firms, I believe in local upskilling. So I really focus on teaching internal teams the ability to build automations like this, know, really focused on tools that they'd have more immediate access to and having them stay there because what you find is you have someone come in, you have someone build these really fancy automations, they worked for three months, your business changes and no one knows how to refine them. So.

Corey Ganim: Right.

Mike Dion: If I build you an automation, I maybe bought you three months of automation and then you have to keep coming back to the well. If I train you how to build these automations, your team can keep this going forever. Right. So that's the really that's really my philosophy is as someone who is a, know, corporate finance insider, not having these consultants coming and going, who's motivated to help your company, your own employees, who knows your company best, your own employees, who's best positioned to automate your own employees. And it's all about enabling people.

Corey Ganim: Right.

Mike Dion: to do this easily and feel the confidence to do so.

Corey Ganim: I see. Yeah. And then that makes a lot more sense too, because I mean, like you said, the businesses change so rapidly, especially ones that are in scaling mode and ones that are growing at a high pace. It's like this might be suitable to the business needs today. But like you said, three months from now, six months from now, 12 months from now, the business might be in an entirely different place. And this, might've outgrown this automation or they've got different systems. Like things are just going to be different. So I think that makes sense as far as looking to upscale the team versus build it once and then hope that they come back for more later. Now, Mike, so this has been fantastic. Where, where do you want people to kind of go and connect with you? Like where can people, if they want to hire you to teach their team, like where do you want to send people after this demonstration?

Mike Dion: Sure, so two big places. So I'd say one of my favorite places is my YouTube channel, F9 Finance, because I like putting out stuff there for free. I wanna make sure that you have access to all this information and it's a great place to get new tools and tech every single week. But aside from that, I've got my website, F9 Finance. You'll find all corporate training there. But the best part is you'll find my free weekly newsletter, Finance Automation Insider. When you sign up for that, I send you 15 five-minute finance automations that you can build with. that you already have on your computer. So that'll save you five to 10 hours a month forever, totally free just for joining the newsletter. And I'll send you stuff just like this, NAVN AICFO every single week.

Corey Ganim: Awesome. Well, Mike, this is good stuff, man. And you did a really good job breaking it down in a way that, mean, like I said, we were 42 minutes in. I mean, it probably took only 30 minutes to build the automation end to end. mean, somebody could just rewind this copy it step by step, apply it to their business and boom, they've got a full working AI agent CFO installed in their business in under 45 minutes. So great work. Thank you for your time. And for everybody that is in the audience and listening to us here, we'll have. Links to Mike and his YouTube channel in the description and in the show notes. And we'll be back next week. Thanks so much for your time, Mike.

Mike Dion: It's been a pleasure. Thanks for having me.