AI in Finance: Connect AI Tools with Confidence

Transcript

Jon Cochrane: [00:00:00] All right. I think we’re live.
Chris Weber: We are live.
Jon Cochrane: That’s exciting. Well, hey, everyone. Welcome to our webinar here. We’re really excited. So for those of you that are joining us, we’re gonna dive right in. Right? So if you’re diving in, jumping on here a few minutes late or if you’re gonna be sad that you missed this opening monologue, don’t worry, you can go back and watch the recording after the fact.
But what we’re gonna do here is we’re gonna have a four-part series where what we try to do is share a little bit about what we’re doing ourselves at Maxio, [00:01:00] what we’re doing within the office of the CFO, with AI, and to keep it practical and to keep it fun for everybody.
If you haven’t met Chris before, Chris actually kicked off a lot of our series and webinars around MCP, so it’s only fitting that he helps us kick off this next series. So Chris, why don’t you give everybody a quick introduction of yourself? I’ll give everybody an introduction of myself, and my background, and then we’ll dive into the topics for today.
Chris Weber: Yeah, happy to. Thanks everybody for taking time out of your day, as Jon was saying, to join us. Pretty excited, pretty passionate about this topic as well. So Chris Weber, heading up product here at Maxio, but my background’s actually more in line with Jon, in the finance and accounting realm.
More focused on operations though, and then joined Maxio quite a while ago and have been really excited to help build the product that my former self would have loved to use and continue to evolve it, and it’s just been such a fun last 24 months, 36 months with the advent and coming of AI and how that’s just rapidly evolved.
Really kind of forcing every role, every function, including the finance [00:02:00] functions, to rethink how they’re operating day in and day out, and it’s been an exciting time to be on the front lines of helping shape that product, as Jon mentioned.
Jon Cochrane: Well, Chris, I’m excited to do this webinar with you.
You know, as we were prepping for this, I shared on LinkedIn a little bit that you and I have, you know, we’ve worked together in a past life. We work together here at Maxio. We’ve looked at a lot of these tools in the past, and so we’re excited to share some of that journey with the group here.
One question I actually have for you, Chris, before we dive into this, you know, as we were… For those of you who are joining us for the first time, we went through a thorough evaluation process back, you know… Well, I would say it this way, taking a step back. You know, we’ve gotten to see a lot of different evolutions of the office of the CFO, different tools out there, things like that. [00:03:00]
And for those of you that don’t know this about Chris, he’s got quite an entrepreneurial mindset. He was there at the outbreak of the iPhone, and I know whenever some of this stuff had first launched, Chris, you said, “Hey, this stuff that’s coming out with ChatGPT, with Claude, with all of that stuff, it’s the most excited I’ve been since the iPhone came onto the market.”
And again, before we get into some of this stuff, and we’re gonna talk about security, how we get comfortable with all that stuff, but maybe just give me one or two of the top reasons why you’ve been so excited with some of the stuff that’s been coming out here.
Chris Weber: Oh, man. It’s a really good question.
I think-
Jon Cochrane: Putting you on the spot.
Chris Weber: No, I mean, it’s a great question. I think all of us should, and it’s kind of funny ’cause everybody goes through the kind of the phases and when you have something that’s kind of foundationally changing and potentially unsettling, the development of like AI and what does that mean for your role and your job.
And I think what I get really excited about once you kind of move [00:04:00] past that is just it’s such an enabler. And maybe it is that entrepreneurial background that you were talking about, but like now kind of everything’s possible, right? Like where you had dependencies and time, like you could automate as many of the kind of time-consuming tasks that take focus away from the why and what truly matters to the ability of like reducing dependencies on skill sets that you might not have.
And I think that’s really kind of exciting is AI is kind of the great equalizer as allowing everybody to have access to information or to explore the curiosity. Like if you wanna learn about something, I mean, it’s just so great ’cause it tailors it to your prompt or your response. So regardless of where you find yourself in your career or your personal life, like there’s just so much opportunity to learn and develop and frankly experiment.
And that does kind of relate to today’s topic of like building confidence of like, yeah, that’s the really exciting piece [00:05:00] that you can learn, you can experiment, but how do you still actually trust and verify? And I think it’s really relevant to today’s audience ’cause I think there’s some skepticism in finance, but I don’t think it’s necessarily skepticism.
I think it’s just, you know, finance and accounting by definition, you have to be accountable to the results. So inherently your bar that you have to meet is substantially higher. But overall, I mean, just things that this time last year took hours or days or weeks now take minutes, and I find myself on the commute interacting with AI and having a conversation with it and prepping so when I land at the office, I’m able to actually get right into the work that materially matters.
So it’s just a super fun time. I would flip it back and curious your answer, Jon.
Jon Cochrane: Yeah. Well, I think I might, I’m gonna hold on my answer there, Chris, because I think we’re gonna get into some of that content here on the webinar. [00:06:00] You know, one of the things… And we’re…
The purpose of this series too, for everybody who’s in attendance, again, we’re gonna do four parts here. We really wanna open up the hood. We don’t wanna bore you with slides, with different things like that. We wanna make it relevant. We want you all to participate in this conversation.
So as we’re going, please ask questions to Chris and I. You know, it’s rare that we get to do a webinar together but we are here because, you know, A, we care about the finance community. We wanna share what our own journey here at Maxio. You know, as Maxio’s CFO, we’re going through our own evolution in how we think about adopting these tools ourselves.
There’s a lot of things to get excited about but what we wanted to do today is to start at confidence and getting comfortable with some of these tools. So some of the stuff that we’re gonna talk about here today is how do you start connecting these tools? How do you start putting them into your workflows?
‘Cause the conversations I’m having with other people who are in my shoes, you know, there is some skepticism and I think rightfully so. Because I know for myself, [00:07:00] just to, by way of kind of introduction to the other folks here on the webinar as well, I’m Jon. I just became Maxio’s CFO a few months ago.
I’ve been at the company now for about six and a half years. My background, I’m a CPA by trade. I started my career in public accounting, auditing, did technical accounting consulting and wore a lot of different hats. And two things that have been pillars of my career but also not only my career but many people who are on this webinar is there are two things that you have to get right from the very beginning.
One is, you know, keep it secret, and keep it safe, the confidentiality piece of it. And the second piece of it is you gotta stand behind the outputs cold, and you gotta know your numbers cold. You gotta be able to have confidence about those. And so I think rightfully so there’s some skepticism within the accounting realm because it’s not that, like, we’re trying to take a step back.
It’s more so, like, we’re accountable to [00:08:00] keeping the most important data confidential and then also making sure that we can stand behind the numbers. And I think what a lot of us are seeing in the market here, and Chris, this kinda gets to your, the previous webinar that we put on here is, you know, there’s some really cool stuff that you can do with tools like MCP, with, you know, with the…
Well, just there’s a lot of stuff going out there. There’s a lot of flexing in the market, a lot of people showing what you can automate and stuff like that. But the barrier between, like, what’s possible and that you’re willing to stamp your name on is, like, having trust within that process. And I think one of the things that, the two things that I think about within here is yes, there’s real risk around making sure that you have things protected from a confidentiality standpoint, and then also, you know, that you have confidence over the outputs that are coming.
But I think there’s the third risk too, and that’s as you start diving [00:09:00] into these tools, you’re gonna see that, you know, Chris, you said that you prep on the way to work every day by just kind of talking back and forth between… I assume you’re talking, you’re not texting to Claude. So-
No.
Yeah. So it’s, like you’ll start seeing that there’s a lot of power in these tools, and we’re gonna open up the hood here in a minute, shift off the slides. But there is a third risk here where it’s that the fear keeps you from leaning into these tools. And so my hope for everybody on this webinar, we’ll show you a couple things today that, you know, before we chat again next week, is that you actually dive in.
If maybe you’re already testing out some of these things, you’re playing with the tools, stuff like that. But my hope is that you actually dive in and start experimenting with these tools because you have kind of two options on how you can get from coast to coast metaphorically here.
You can ride a bike, which I like biking, but it’s gonna take me a long time to get [00:10:00] from coast to coast, or you can hop on the plane and fly. Chris, one thing I, you know, as we were talking about, just the mindset that you have to shift into here, you brought up this concept of being the manager over these AI tools.
So I’m curious if you can just expand on some of the stuff that you were sharing with me, not only in how you’re teaching our product and development teams to deal with AI, but just more broadly, like if you’re gonna use these tools, like how you have to adopt the managerial mindset.
Chris Weber: Yeah. So a couple things, and on your last slide too, I think the most important thing, what you were talking about is taking that first step. And we’re not talking… I think everybody joining this webinar, and we’re talking a little bit more generic here, ’cause I think everybody’s at a different phase of that journey, right?
It could be you’re just starting to experiment and use more of ChatGPT or Claude in the true basic level. Maybe you’re… Like in past, we’ve surveyed [00:11:00] the audience, likely some of you have attended those, and it was kind of mixed results of where people are in that journey.
Others are dabbling a little bit in the connectors and MCPs. Others are starting to leverage skills and automations. Others are going full bore with like a Claude code and actually building micro kind of applications to serve these needs. Regardless of where you are on that journey, I think to Jon’s point, it’s like taking that first step and focusing on what are the next three feet of the world.
Not necessarily always how far in advance somebody is or your perception of being behind. It’s just being on that journey. Now, Jon’s question was more about the mindset of the manager. I think some of the hesitancy of leveraging AI is definitely the confidence that we’re talking about, the being accountable to the results.
I think a great frame of mind that I’ve been encouraging everybody that I work with to think about is you’re now kind of the manager and you’re delegating tasks to AI. Right? So just as if you had hired somebody [00:12:00] new on your team, a new FP&A person, like you’re not gonna in week one throw them the full board deck and full management package and MD&A and say, “Hey, generate this and ship it directly to the board,” right?
That’s trust and confidence that’s earned. You gotta like gradually work them through that process. Trust but verify, right? And I think the important part of this is find what are those time-consuming tasks that have very low level of risk, and automate and start there and build your confidence. And why I’m again being generic is like that could be just drafting Slack messages or emails.
It could be building Excel formulas for some analysis you’re doing. It could be building small skills, but you kinda gotta balance that risk/reward, just like as you’re training somebody that you’re mentoring, what are the tasks that you would delegate, and then how do you continue to review that and coach it?
And I think where people get it wrong is, just as if you’d hired a new employee and then never kinda coached or verified or [00:13:00] leaned in to teach them, you can get the same similar results with AI. So thinking of it in terms of, you’re not just managing people, you’re also managing agents or AI or these automations that you’re building, I think is a really critical frame of mind.
Jon Cochrane: Well, I think that’s super helpful, Chris.
And yeah, I think this is a good point for us to start jumping into some of the meat and potatoes here, and actually sharing a little bit around… You know, and by the way, for the folks on the webinar here, like I’m pulling up my actual Claude instance. We’re gonna show you some of the things that I look at to try to get comfortable myself within how I’m leveraging it here at Maxio.
And then Chris also has his own Claude instance and a ChatGPT Codex instance up. So this is also the time, like if you have questions, you want us to elaborate on certain things. There’s… We have a loose agenda here, but this is also for you. [00:14:00] So how do I trust that it’s safe? So this is like a big, a very big thing, especially when we think about finance.
How do we make sure that as I’m prompting Claude, that data’s not going where I don’t want it to go? And so the first thing I’ll say before I open up Claude, I think this is probably well known at this point, but I don’t wanna assume anything. You wanna make sure that you’re using a paid account.
There are, you know, as I think everybody can see, let me make sure. Well, that’s my old browser. Let me get to the right window here. And we’ll go entire screen, screen one. There we go. All right, here we go. So actually I’m in a project here because we’ll dive into some of the projects I was trying to set up for this webinar here.
Some of the stuff that, when we think about different settings within your Claude account, so you should be using, all of us should be using a paid account because if you’re using a free [00:15:00] account, these are the accounts where like the stuff that you enter can be used to train AI, right?
Last thing you want is your company’s financials to go into any sort of training data, and that’s one of the big things of shifting from a free account to a paid account, whether it’s a personal paid account or a teams account. Chris, I’m gonna pause there, because I’m sure you have thoughts and opinions on just that first level there.
But anything else that you would add?
Chris Weber: Yeah. I mean, it’s really good practical advice too. You know, I encourage every company to make sure that you have kind of an acceptable use AI policy. You know, fully acknowledging some companies might be too immature to have that, but, like, I mean, just the standard advice of having a basic paid account, you know, strict access, limit the access.
Jon showing some of the configuration in this instance. Might have also caught we had two different organizations there, right? Separating kind of go-to-market to engineering, ’cause we needed a little bit more granularity of overall organizational permissions. So, like some [00:16:00] techniques that might increase costs, but, you know, Claude and OpenAI continue, or Anthropic continue to kind of get more granular in the permissions.
But, you know, like many people on this journey, depending on where you’re at, just get started, experiment. You know, if your company is in a more sensitive environment that doesn’t permit the use of AI, right? Some of the techniques that we’re talking about, just take out of the context of finance and apply to your personal life ’cause it is so important to still be experimenting and iterating there.
And in fact, like you could take some of these concepts that Jon’s showing and create a paid Claude account and do some of your personal finances, you still want the same security around. So yeah, I think you nailed it, Jon.
Jon Cochrane: Yeah. So here’s, like a practical thing right now. So right now in Maxio, you know, like to Chris’s point, we do, you know, we have one instance that we set up for our developers and engineers ’cause they have different settings than what we need from, from like our finance go-to-market type [00:17:00] organization.
Some people even have their own instance for finance because you just wanna make sure that you have an extra layer of security there. But one of the things that you’ll notice here is I’m in a project right now. This project was created by me, and you can see that this project is private.
So let’s just say that I’m getting started, and one of the big reasons why you wanna actually operate within a project is context and storing that context. And we’ll get into this in a little bit here, but especially when we get into trusting the outputs of a system. But one of the things that you don’t wanna have to do is explain the same thing over and over to Claude again.
So one of the things that I like to do is I have a project, and within that project I have settings. So I could go in here, I could create a new project, and what it’ll ask me is, “What are you working on? What are you trying to achieve?” And then because I’m in a Teams account, this is where I can choose the privacy of that project.
Do I want it to just be available to me, or do I want it to be available to everybody at Maxio? Now, one of the really cool things is if you are working on certain things, maybe you have a [00:18:00] corporate, you know, like a certain part of your project that you want… You don’t want everybody to have to go and create, and you will have a central repository that everybody can access.
This is like a really good way to collaborate across a team account. But you know, I’m dealing in finance stuff. I don’t need anybody to know that stuff except who needs to know that. So most of my stuff is gonna be private. Some of the other stuff that I think about too, so I’ll go back into my project here.
But when you think about what you’re actually connecting to Claude itself, so let’s just… You know, there’s a lot of different settings here on the left. You know, I was clicking around a little bit while Chris was talking there. But when we think about connectors, you can actually go to the Claude store.
I can go browse connectors, and there’s a lot of different things, and there are new companies getting added here every single day. But it’s like, how do you know who you’re connecting to? How do you know what permissions you’re giving those tools or things like that? Well, you can actually control this stuff.
So I’ll go back real quick, and I’ll [00:19:00] go to, let’s just say I have my Gmail connected to here. And so within Gmail, I think this is an area that a lot of finance professionals are getting… You know, it’s like a really, it’s a low pressure way for you to kinda start with this stuff because you’re not dealing with financial data, but you’re starting to build some capabilities here.
So I have different things between, I have read-only tools, I have write/delete tools. And I think that’s all I have within this Gmail connector. But you can see here I have settings within there. My read-only tools, I’m fine if Gmail reads all of my stuff that comes through. So I’m always gonna allow that.
But the moment that you wanna write or delete anything, you gotta talk to me first. Or maybe I wanna set that to never, never allow. These are, like, certain things that you can start plugging in and setting up boundaries for, like, how do you keep stuff safe, secure? How do you prevent stuff from taking actions that you don’t want it to take?
Really, really helpful measures here, [00:20:00] just from different connectors, and this applies to different things like that. So for example, I have a NetSuite connector here, where I have a lot of stuff set to need approval now. But when it comes to getting different things or listing my accounts or the write/delete tools, stuff like that, everything’s kinda set to needs approval.
Because now all of a sudden I’m starting to get into the ERP space or the GL space, different things like that. Chris, anything that you would add so far to what we’re talking about here?
Chris Weber: No, I mean, I think the thing to keep in mind is that you can adjust these permissions as you go. So you could start out and be pretty restrictive and have it ask for approval for everything, and then figure out what kind of matches your use cases or your needs and then adjust as appropriate.
But what is nice is that with the MCP, you’re generally authenticating. So like the Maxio MCP, you’re authenticating through your credentials that you have. So just because your [00:21:00] organization has enabled an MCP connector here that’s available, each individual user still needs to authenticate with their own authentication and login.
So I think generally I’ve referenced this before, I think usually it’s helpful to think of it in terms of an API, right? ‘Cause it’s something that’s a little bit more tangible to most, but it’s a little different than an API key. It’s not just an API key that could potentially be shared, right?
You’re connecting it, it’s a secure connection, and then you’re still authenticating through, and then each individual user can kind of change these settings. Like Anthropic or Claude in this case has the ability as a organizational admin to restrict certain tools holistically, which is a nice benefit, and then each individual user can go more granular on it.
I think the other thing too is Jon was talking about like projects. You can enable certain tools within the project, give it custom instructions. So there’s a lot of ways that you can kind of restrict this and build that confidence over time. And I [00:22:00] think going back to the reference Jon made earlier about being the manager, right?
Think of AI, in this case Claude, as that new employee. If you’re trying to build that confidence and trust, start gradually and then give it more and more responsibility as you see suitable.
Jon Cochrane: Yeah. Super helpful there, Chris. Well, let me switch over here because we’ve talked a little bit about, you know, one of the things that Chris and I, we’re not gonna dive into here.
You know, Chris and I are not IT professionals. We are not, like, system experts or things like that. But I think these are some of the securities and procedures that you can actually set up today that can allow you to start experimenting with this stuff in a safe and secure environment.
One of the things that I wanna cut over to next is some of the… A little bit, so we talked a little bit about, you know, can I trust that it’s safe? Chris talked about starting small, like maybe we just [00:23:00] start with some read capabilities before we actually get into acting within the data.
We showed a little bit here within Claude how you can start setting up security procedures. Or sorry, not security. You can set the permissions for what you want to allow it to read or prompt or not allow you, or to act automatically on. I think it would be helpful to start diving into: Can I trust what it gives back?
And so this is where we actually start leveraging some of the capabilities that we have within Claude. Or, you know, whether you’re using Codex or one of these other AI tools. Like how can you trust… You know, you hear all the time about how AI will give, confidently give you answers, and it’ll like lie straight to your face.
Well, how do we get beyond that and start developing some stuff where you can start trusting the outputs? Because once we start trusting the outputs, then all of a sudden, like we’re stepping off the bicycle, we’re boarding the plane, and we’re about to go really fast because then it gets really fun.
[00:24:00] Chris, I’ve got a couple things queued up here on like a basic level. But if there are anything that you think would be helpful from some of the demo environments that you have set up, also feel free to jump in here at any time to share some of that stuff. I’m gonna set up one little…
I’m gonna flip back and forth between two different… I’m gonna shift my screen share around a little bit just to show a few items.
All right. And again, as a reminder, if you have any questions, items, things you’d like us to demo or show that might be helpful, please feel free to jump into the chat.
So
Let’s go ahead here. I’m gonna share my entire screen. And now you can actually [00:25:00] see two things here. So I’ve got my Claude set up on the right. I’ve got a sandbox set up on the left. And one of the stuff, I’ve already gone ahead and actually connected my Maxio instance to Claude here. So I wanna start showing some examples of how I can get comfortable with some of the outputs.
And first thing I did here, perhaps some of you are Maxio users today, perhaps some of you are not Maxio users yet. This concept holds true whether you’re using Maxio, QuickBooks, NetSuite, Campfire, Relt, like you name it, Intact. Like, this type of procedure would be something I’d recommend that you start with from the very beginning, otherwise you’re never gonna get comfortable with it.
So within here, I went into the admin, and within the integration section, we have a Maxio MCP here. And you can see I actually have two different connectors actively set up here. So this means that, to Chris’s point, two different people went in separately, authenticated, and, like, their per-user authentication is set up [00:26:00] here or is connected to this instance.
So at any point, I can click in here, and I can see, hey, somebody connected, created a connection on October 31st. And then there’s another one. Whoa, I’ll X out of that. And then another one created one yesterday, and that was yours truly, as I was preparing for this webinar.
So within the MCP here, within Maxio, I can actually see, okay, what are the different operations? There are no create operations here. There are a few read operations created. And then no update operations are connected. And now that is because of how I actually set up the connector itself. So within, this is within, this is Maxio’s MCP.
This is what we’ll kinda demo today. But again, this applies to, like, QuickBooks, et cetera. So there are two different roles here where one is a read-only role and another is a bookkeeper where you can see this has, like, read and write capabilities. And so when you’re thinking about connecting [00:27:00] your data, like, think about, okay, where do I wanna start?
Do I wanna start at the analyst, the read-only, where it can’t take any actions, or do I wanna go to the read and write, where this has nine write scopes? We’re gonna demo the read-only for today. And before I go any further, I do see a question in the chat from Mina, and it says, “Do you think we can use Gemini with Gmail to reprioritize and escalate if no action is taken within twenty-four hours?”
Chris, I’m gonna throw that your way to get your thoughts.
Chris Weber: I love that one. Yeah, so Gemini is really interesting. I personally don’t have as much experience with Gemini outside of it being embedded in Gmail. I mean, we’re a big Google shop here at Maxio, though. I think that the philosophy of what you’re talking about is exactly right.
And I think, too, the hope is you can find ways to build these automations, so you can build scheduled automations and tasks. I actually have a few that are very similar to what Mina was asking, and Jon was [00:28:00] referencing my drive-in, right? Like, every morning I have it scan any of the emails from yesterday and provide a summary of those and help prioritize the action plan.
Because I’m now focused on product and engineering, I do the same thing. I have a summary skill, and this is one of Anthropic’s default skills, actually. You can connect it to GitHub and summarize any new pull requests or changes. And those that aren’t familiar with it, that’s more of, like, changes in the code or progress on certain updates.
That’s a phenomenal use case of AI. And I think we’re gonna get to it a little bit later on as well. But some of those use cases when we’re talking about, like, low risk, that’s a really good example of low risk ’cause you’re not yet delegating it to propose actions. You’re asking it to summarize, “Hey, what emails did I miss?
What do I need to bumble back up to the top of my inbox?” And then maybe the next piece is you’re talking about recommending, [00:29:00] right, to tie it into Jon’s slide. Maybe, Mina, then you take it one step further once you’re comfortable with that and say, “Hey, go ahead and draft some email replies.” So then in the morning I’m coming in and I have in my draft inbox replies, so I can edit it.
And then maybe eventually you get a high degree enough of confidence that you start having it act and send those. But those are different techniques that you can use, but I think that’s a very relevant one.
Jon Cochrane: Yeah. And one thing that’s pretty cool, whether you’re using Claude, Gemini, OpenAI, Codex, whatever it is, I mean, these AI companies are in a race with each other.
So a lot of what we’re talking about here from, like, what you can do with one, you can do with another. So I think it’s good to just dive in. So now one of the things that I kind of kicked off here just to show you that, hey, a lot of this stuff is live. I have some stuff queued up as well just to show you a way that you can kind of get started on a lot of [00:30:00] this stuff.
But one of the… Well, let me take a step back here really quick. I wanted to show that, hey, we actually have a query that’s running right now. I’ll come back to that. But one of the things I did last night as I was preparing for this demo is I, A, just established the MCP connection, and then I said, “Hey, make sure that you’re connected and then tell me what you can do.”
And then it walked through and told me, “Hey, I’m connected, and here’s a couple things that I can do.” So this is leveraging our advanced billing and core capabilities. What this can do, this can pull details about our customers and contracts, our billing and ARR, different details about our subscriptions, reports, projects, a lot of different things like that.
And then I actually have a demo customer here, and on my left here, I’ll flip back and forth because one of the things that Chris talked about earlier is, like, how do you start getting comfortable that the data that it’s giving you is accurate? And just like you had a new… So just [00:31:00] like you’re the manager now and you have a new hire, one of the things I encourage you to do is start with a single record, and then query, ask it questions just like you would, you know…
I think about a lot of people that I have trained over my years on different teams and stuff like that, and you think about, like, a first-year hire or a new hire on your team. How do you trust that the information they’re giving you is accurate? Well, you ask it questions that you know the answer to, and then you make sure that it’s kind of stepping through the processes.
So for example here, I said, “Hey, could you tell me some details about Affirm?” It’s kind of an open-ended question. Open-ended questions aren’t always the best way to deal with AI, but hey, why not? So it said, “Hey, here’s some details about Affirm. The customer ID is four three seven six.” I can go into here, and I can actually see, oh, hey, the customer ID for this is four three seven six.
I can actually see that in the URL up here. The number is this. Okay, we’re on a roll here. And then it can tell us any details about the contracts with it, the different contract line items within that. It’s like you have three [00:32:00] transactions. That’s very cool. I do have three transactions.
Here are the amounts. Here are the start date, end dates. Here’s the period duration. And then actually this is where sometimes open-ended questions can be helpful, where it says, like, hey, actually not only do I see these three line items, actually notice that this customer expanded, went from forty-three thousand to forty-seven thousand three hundred, a ten percent expansion year over year.
And here’s a little bit more details. Then I was like, “Okay, that’s helpful.” And you can start stepping through these things piece by piece by piece as you start getting comfortable with how does it extract the outputs? How do I put the inputs into here? Different things like that. Here’s Vera actually had a question for us in the chat.
Chris, I’m gonna throw this one your way as well. Maybe I’ll just make you answer all the questions here.
Chris Weber: Yeah, happy to. [00:33:00]
Jon Cochrane: Here, and I’ll ask it out loud for everybody. So Vera said, like, “Hey, I’m worried about uploading too much information, but I wanna use AI to generate different revenue forecast scenarios, factoring renewals, pricing new products.”
So I mean, Chris, where do you wanna start with… There are a lot of pieces to that one. We’ve actually gone through some of this capability, some of this ourselves here at Maxio, but curious where you would wanna start with that one.
Chris Weber: Yeah, I think there’s two aspects to that question.
And it’s similar to Mina’s in the sense of more exploratory, right? Like, I wanna get started, but I have these concerns. So I might actually wanna share my screen here in a second, Jon, ’cause I wanna show a technique that I think sometimes we all tend to just gloss over ’cause it’s a pattern, a thought pattern we’re not as used to.
And then there’s the other aspect to that question of like, is that a good use case? Is that the best use case? You know, I think we’ll get into the question around deterministic [00:34:00] outcomes versus more analytical or analysis outcomes. So let me try to answer both. But the first thing I wanna do is share my screen here and let’s take a look at this.
Looks like you guys can see it. So I actually just loaded Mina’s question in, and this is a technique that I wanted to show because sometimes you just need to get in and have a conversation with Claude or with Gemini, right? So in this case, I’m in Claude. It’s obviously gonna be biased towards Claude.
It’s suggesting what Gemini could probably do, but then down here is saying, “Hey, this sounds like something I could do. Should I do it?” Right? And this is the thing that I think is we tend to gloss over, is you don’t have to know how to do something on the outside.
Jon Cochrane: Yeah, it’s powerful.
Chris Weber: We’re so used to having to go and figure out, how do I go do this thing and then get started to do it?
Here, it’s like, “I don’t know. Here’s what I’m thinking about. Help walk me through it,” right? And it’ll actually… You [00:35:00] just have to have a conversation. And this is how so many of these, like, micro applications get built. You start with a problem, and it’s like, well, here’s what I potentially could do, right?
And it’s starting to look at what you could build, and it’s referencing these external documents. And then it’s gonna start weighing through, okay, how do I actually build this? At a certain point, it’s probably gonna prompt me and say, “Hey, do you want me to go ahead and build a skill? Do you want me to go ahead and schedule this?”
You can really actually use AI to walk you through and educate you on how to do something. Now the tie-in to Vera’s question is something that I would phrase around where you see AI getting things wrong, right? And so here’s the crawl, read and prioritize, schedule a cohort task. Great.
Walk. Right. Still, probably, yeah, the daily digest, the run, autonomous escalations, and what actions do you wanna take? So you can see where it’s going with it. And what’s really nice with [00:36:00] Claude that I enjoy is it’s gonna ask me a series of questions. What counts as an action taken, right? And this is a great kind of follow-on question ’cause Mina was asking, like, an action taken.
Does that mean that you’ve archived it? Is it still in your inbox if you’ve replied? So it’s really asking clarifying questions, and you can see the other questions of which inbox is this for, right? So you’re already starting to see how this is gonna develop. You know, let me stop the screen share and go back to the other piece on Vera’s…
There we go. You know, there’s a question around deterministic logic, and I think a lot of the pushback on AI is like, “Hey, let me just dump…” But for example, like, “Let me export all this data from Maxio and have it generate a revenue schedule.” I think the challenge with using like an LLM in that aspect is if you need the same answer a hundred times out of a hundred times, an LLM’s probably not the best use case.
So how do you frame it differently and say, “Hey, how [00:37:00] do I actually write a SQL query?” Or, “How do I write an Excel formula?” Or, “How do I build a mini application that I could run locally to do this task,” right? Now you’re using AI to build deterministic logic. Now, the key in what I’m stating, though, is, like any problem, there’s probably four solutions.
As you all probably know, to make it relevant in Excel, you could do a V lookup, an X lookup, you could do if statements, you could do hard codes. There’s so many ways to build a formula in Excel. The… Multiple of them could be right, some of them could be more efficient, right? That’s the discretion that you’re giving to the LLM to build something that is then deterministic that’s gonna give you the same answer a hundred times out of a hundred times.
So that’s the aspect of like, what do I use? You can also use AI to help build the SQL queries or terminal or make it more digestible. So a couple techniques there in answering that question, but great questions.
Jon Cochrane: Yeah. In line with the [00:38:00] security piece of it too, Maurice had a question here that I think is really good, and important for us to chat about, and I’m happy to take…
dive into this one a little bit here. So it says, “I have one question. How secure is the AI tool when it comes to storing large amounts of information? Specifically, what measures are in place to protect that data from unauthorized access or external security breaches?” So one piece of, when it comes to unauthorized access, external security breaches, stuff like that, I think this is where, like, having a really solid relationship with IT comes into place.
Like, you would think about, “How do I think about protecting my company’s data in general? How can people not get, you know… How do I make sure that nobody can, like, get into my email?” Finance is always a target of phishing attempts, stuff like that, people trying to get into the system because you have the keys to the kingdom.
So partnering with IT, important thing in and of itself. When it comes to how do you actually access, like where does the [00:39:00] data live? How do you access this data? How do you make sure that Claude doesn’t take the data, push it somewhere else? I think this is where some of those security settings, or not security, like the permissions that you give Claude is important.
You can actually set organization settings within Claude as to, like, you can’t access certain websites or stuff like that. But more importantly, one thing that I think, and this is where you kind of get into hallucination a little bit, or you can get into hallucination. One of the key things, like, getting the most out of AI is figuring out, like, where…
Like, how do you make sure that Claude or Gemini or ChatGPT or whatever it is, is pulling from a source of data that you trust? And where does that source of data live? So you’ll… Let me share my screen again. One of the things, one of the main things that I do here is at least when I’m accessing data, whether it be from Maxio or maybe I’m pulling it from [00:40:00] QuickBooks or an ERP or something like that, the data lives in my system.
And where I spend a lot of my effort is making sure that I, like say Claude is like my, like a new hire on my team. I’m training Claude. I’m like, “Hey, when you go pull data from here, here’s how I want you to interpret the data,” but the data always lives in the sandbox. Once you start pumping a lot of information to Claude, and if you’re trying to use Claude, especially like a project as a, as like a database or a warehouse, and you start pumping it full of a lot of information, you will…
This is where you’ll find that sometimes, like you’ll ask it, “Hey, what’s two plus two?” And it will say like, “It’s six.” And you ask it again, and it’s like, “It’s four.” Then you ask it again, and it’s like, “It’s eight.” But if you know that like within your system it’s, well, what are my total transactions? You know, like I could actually just ask this here, and say, “Give me [00:41:00] the financial breakdown of Affirm.”
I’ll see what it comes back with. And then we’ll see what this comes back with. I know that at the end of the day, hopefully it returns these values to me, but the data lives here.
Chris Weber: Yeah, and I think to-
Chris, go ahead.
To net that out, right, you, it was kind of the prior point I was making as well is you gotta be, use some discretion on what data set that you give it, and it does tie into the deterministic logic as well.
Like, you’re relying on potentially Maxio, the ERP, as the system of record and the source data where your balance sheets can always balance in your general ledger, right? Like, these concepts and protections that Maxio has on this customer or Affirm to ensure that your transactions reconciles to the revenue, to the invoice [00:42:00] schedule as well, right?
You’re not asking it to recalculate all that. You’re presenting that and bubbling that up and surfacing insights. So if Jon were to scroll up in the thread, right, like he asked it to pull some of the reporting data, and it’s showing the 17 customers that were expired, up I think it maybe it was a separate thread-
Oh, it’s like a different…
Yeah… from the one you had. Yeah. One you had on the thread, right? It’s saying, “Hey, this is abnormal for your account. Here’s customers that have contracts that are expiring. Here’s their dates, here’s the tier of the plan that they’re on, and here’s what the ARR,” right? That’s not deterministic, right?
Maybe next month there are no customers that renew in the month of August, and so it’s not gonna surface that same insight. That is the kind of generic version of, like, use your source of truth to do those core calculations and then surface the insights, the things that you would be, again, going back to my reference of a manager, that you might…
And even if you’re, like, just starting your career in an individual contributor role, there’s no reason not to have [00:43:00] that manager mindset of, like, what are the insights that ideally you’d have somebody else on your team surfacing to you, and that’s a great way to surface ideas for AI. Part of it too is, and Jon, maybe I can share my screen here to show-
Yep.
a different example. Part of it’s on your providers, right? And Maxio potentially being one of those. I know we have a mix of prospects and customers on this webinar, but let me show you a couple other things. And it’s something that we take in to heart as we develop the product as well. So let me make sure that my connectors are enabled here, Weber & Co.
Great. So this is similar, just building on the example that Jon was doing, pulling the information for a customer, in this case, a subscription, right? And the ability, Jon, I love the dual screen, right? Validate the information that’s getting returned, but part of it is on the provider of the connector to ensure that those things that are a little bit more tricky per se, [00:44:00] is surfaced in a way that you can trust and validate and verify.
So one example would be, like this. Let’s say, let’s increase this subscription by one seat, moving from zero user seats to one, right? So this is a good example of Maxio realizing, “Hey, what’s the risk profile of this change?” And it’s, “Hey, we’re gonna actually adjust this subscription and move the quantity assigned from zero seats to one seat.”
Maxio is actually gonna return an embedded UI here demonstrating, hey, a preview of what that change looks like. Here’s the amount that’s due now. Do you wanna change it and maybe move it to a quantity of two, right? Preview that change, and then actually confirm and submit that change. So some of it is also on your providers of these connectors.
I think [00:45:00] Maxio is kind of a little bit forward than most connectors at this point, but this is ultimately where most providers will start going, is having its own risk profile of the changes and surfacing their UI like Maxio does here to affirm and confirm the change. And then I can actually go ahead and confirm and make that change, and then I could jump over, log into Maxio and see that that change has been made.
So almost embedding the Maxio UI inside of Claude or OpenAI, like meeting you where you’re at, so that you can fundamentally do that. So I wanna say, you know, balance this conversation. I think it’s a nice perspective having me on here as I can balance it. It’s both a responsibility of the users to validate and grow that confidence.
It’s also a responsibility of the providers like Maxio to ensure and put those safety guardrails on there. Another thing that like Maxio does, if you’re using the Maxio MCP connector, it logs it as an audit trail as if a user had performed those actions. So if you’re running a report and Jon were to navigate into Maxio, [00:46:00] you’re gonna see that that report was run, right?
You’ll see the timestamp that it was run. If you log in and view that change that I just made for that component change, you’re gonna see the audit trail and the timestamps and the modification get updated as well. So making sure that there’s still that tie back and that it’s not performing actions that are invisible.
And I think maybe it’s a bias towards that concept of being a manager, but if you had delegated that change to somebody on your team of, “Hey, this customer Sunglass Hut needs to go from zero to two,” right? “Go perform that action,” how do you trust but verify? You could easily log into the Maxio UI and see that that change was performed.
Jon Cochrane: Super helpful. I hope some of you are starting to see some of the power that’s behind these tools and as you start getting comfortable with the data, what’s possible. So I kinda wanted to show… I wanna jump into, [00:47:00] if we asked another example here where it’s like how do you validate AI finance output or analysis?
How do you provide to management or stakeholders given the outputs can always change as a result of tweaks in your prompt? This is really good, and honestly, this probably… This opens up kind of, like, where we are, where I wanted to end at least series one for this week. Because the goal of this four-part series is to, you know, we’re starting with basics today because, like, you’re never gonna get to phase two if you can’t get comfortable with, like, how do I start experimenting with this?
And I think Chris gave one of the ways that I’ve learned this stuff myself is starting to ask AI what it can do and then continuing to build off of it. Like, there’s a lot that you can do with this, and you don’t have to figure it out. You don’t have to know it ahead of time.
You can just ask it. It’s like you have an advisor right in front of you that can talk to you and coach you through it. That is a great way to start here. So but let’s get into into the example here where it’s like how can you actually start getting comfortable, [00:48:00] with the financial outputs if there’s a risk that the output’s gonna be different every time you ask it?
Really great question, and this is where we get into how do you start saving context, and how do you start teaching it, and how do you start making sure that when you prompt the data you get the same response every time? So the first piece of that, what I have over here, again, this is… Whoa. Let me share the whole screen, not just one page.
Stop screen share. Let’s get the whole thing. Share screen. Let’s go entire screen. Here we go. Okay, so what I have over here to the left is an ARR summary. I’m comfortable with the first starting point again, is I’m comfortable with the underlying data source that I’m prompting from. And I think one of the hardest parts within finance– Well, A, if you’re comfortable with the underlying data source, that’s like table stakes.
You gotta make sure that you feel… So whether that be your spreadsheet, whether that be QuickBooks, whether that be your ERP, in this case it’s Maxio. Like, if you– That’s like your [00:49:00] starting point. Where is the data? Like, where are you pulling the data from? The analysis on top of that, like, I spend countless hours trying to prepare for board meetings, stuff like that, to unlock the analysis on top of this data because there’s just a lot of different ways to look at this.
And then also, we haven’t even gotten into connecting multiple data sources and pulling them in so that you have context across different data sources. But let’s start here. So I started from a blank report, and I just said, “Hey, are you able to review…” Like, I ran this report over here and I said, “Are you able to view that data?”
And it said, “Yeah, I can see the data, and here’s the data I see. In July 2025, the opening value was 19.75 million.” Yep, check that there. And I can see the ending value, 20.15 million there. And then here’s June and June. And it’s like, all right, great. I have the data there. Well, that’s accurate, but, like, is it helpful?
Not really. And so what I did is, let’s just say I was preparing for a board meeting. First of all, I’m glad that this is comfortable here, but I need to go a layer deeper on this. [00:50:00] And what– As you start having these conversations with Claude, especially in these long-running chats, at the very end of this, you can ask it a question and be like, “What did you learn?
What are some of the takeaways? Can you save this context in a skill that I can then come back to and build off of down the road?” So one of the other things I did here, just to show what’s kind of possible, is I took this report, I modified it slightly. I added a bunch more fields to it. I added customer number, stuff like that.
I added net dollar retention down to it. And so it split, sliced, and diced this data a lot of different ways. So now not only does it show me some of my ending values here, the 27.5, but it gives me like expired transactions, different things like that between active and expired. And I have a couple settings.
You know, I didn’t go through all the settings on this report, so bear with me on that. But I said, “Okay, help me pull some more data using this new run ID that I have up here, and then use this skill to interpret the data and-” [00:51:00]
All right. I’m really bad at typing and talking at the same time. So what I did is I said, “All right.” I sliced and diced this data a few different new ways, and I said, “Use this run ID, and then use this skill that I had created.” And so one of the things that I did to really dive into your questioning here was, you know, I’ve done a bunch of different analyses on the back end, and as I had worked with Claude, I’d said, “Hey, every time that you run this report, this is how I need you to interpret the data, and here’s what the different data pieces mean, and here’s why those are important.”
And so you can actually click in to this skill, and what this basically does is this breaks out, okay, here are some of the different… One thing I didn’t actually check before I ran this report was, you know, I actually have a very specific report I use when I’m running [00:52:00] analyses here at Maxio.
And I said, “Okay, these are the things that matter to me.” I care about whether I’m looking at the header level or the detailed level. I care about the category. I care about customer numbers. I care about the business line, their segment, their start month, start date, start month, and then ultimately when they’re renewing, and these are different fields that we save in our own Maxio account.
And then I care about, I provide debt credit, provide a definition such as, here’s opening loss to ending expand, different things like that. And then this helps interpret a lot of the data within this. And here are some of the mechanics that matter, like what does an expired customer mean?
What does a customer who leaves and then comes back to us mean? These are all… And by the way, these are things I did not type into here. These are things I pulled directly from Maxio’s documentation and said… But again, Claude is a different company than Maxio, so I had to teach Claude. I was like, “When you’re pulling data from here, here’s the support article.
Take that support article, drop it into the chat, and help [00:53:00] develop a skill.” And the skill is something that you can ask Claude to create, and then what you can do is you can review the actual skill itself and then validate this, refine it over time. So these are some things that you can do to make sure that you have the same output every single time that you’re actually running these prompts, leveraging MCPs, pulling data together, making sure that the stuff that you’re using is accurate.
Chris Weber: Yeah, and Jon, one thing I’d add too is, like, that judgment-based kind of work is also one of the advantages of AI too, to kind of spin it like a little bit differently is like, yes, everything Jon said, absolutely, to make it more deterministic versus probable or judgment based.
I actually would say, like, I bet if we gave this audience the exact same data set and we said, “Hey, you can’t use AI,” we would probably all come back with different interpretations of the data. Yeah. That’s actually one of the great benefits. And that’s where, [00:54:00] like, I actually view that judgment-based work one of the benefits of AI, ’cause sometimes, like, you’re crafting a narrative.
Like, if you’re doing your management discussion and analysis, your board reports, your readout, you’re coming into that conversation with some biases. And so having AI kind of review that and challenge you and provide a different perspective can actually be really beneficial. And going back to that example is, is like a manager, right, or a leader, if you had somebody in FP&A put together the outputs. Just like you would, you’d be like, “Yeah, that’s a good point, but let’s go a little bit deeper, or did we consider this?”
Those are the aspects that you can kind of go back and forth with AI, and it’s truly a force multiplier. And you can also get it… Like, I still sometimes use the personas, right? Like, when you get that report, it’s like, “Okay, now put yourself in the shoes of the board member. What are the hard questions that they’re gonna ask?
And help me prepare those and surface those.” So you can actually use that judgment-based aspect of AI to your advantage. But that’s why I do go [00:55:00] back to, like, making sure where it needs to be deterministic, it’s deterministic. Ideally, you relegate some of that to your systems of truth, like your Ps or your GLs.
So there’s a few techniques that you can use. One of the questions that came through was, we’ve used the conversation of skill a lot. Those actually are more of a, starting to become more global. You can basically build skills which are repeatable things that you can call inside of AI.
One technique that I actually wanna show here before we wrap up is actually, you can also build things that are less intuitive, that are helpful. So I’m gonna share my screen. All right. So one thing that I do is, like, it ties to the question of how do you validate things.
I do some hotkeys, right? So I leverage a Mac. You can do text replacement. So I use this all the time, [00:56:00] right? So that prompt is not a skill. It’s not saved in Claude. If I was in email, I could do the same thing. And I’ll pull it over. I think you guys can see it, right? You can actually come in and build these text replacements, and I’ve got a whole host of them, where I get tired of kind of asking the same thing and then frankly, I start to get a little bit lazy of like, hey, maybe I start to be like, “Well, validate that answer,” right?
Or prove out the calculations right. Now, I just have this hotkey of validate the answer you just provided, right? Treat it as if you’re reviewing it from a finance perspective. Identify those assumptions, right? That’s a quick way to get repetition in the way that you’re asking the question before you start to actually build skills which you can come in and build custom skills.
This is a demo environment, so I’ve got a few. You can actually browse skills, and this is the other thing that you’re gonna see more and more providers doing, Maxio included, is actually providing a set of skills that have that instructions and [00:57:00] those automations, so you can build that in a repeatable fashion.
So when we’re talking about skills, it’s truly more of a technical aspect of, like, an actual tool and utility within, within AI.
Jon Cochrane: Yeah, and we’re only a few minutes from our time here. So one of the things I wanted to close the loop on was that actual prompt that I just ran here.
Let me show that back up. And then Anthony, I see your question here in the chat as well. So we’ll answer this while we’re pulling that up. So if you… Just quick refresher here, we started here. We had a report. We said, “Hey, our ending ARR is twenty-seven point five million.”
But then we reran the report with some new attributes. Our ending ARR actually here on the left is now twenty-nine point nine. There’s a good reason why it’s that, and we’ll get into that here in a second. So similar type prompt here. I started with something open-ended, and I said, “Hey, pull the data here.
Now run the same data [00:58:00] with new attributes using this skill.” What it did was it ran this stuff here, and it said, “Hey, validation is working. This is all great stuff, but the first thing I noticed, your reports don’t equal each other.” And the main problem here is that you have this setting turned on.
It’s called carry forward expired transactions. So what this means is, here’s something. If you remember, I said, “Give me some things that I should probably talk to our board about.” And it’s like, “Hey, of your twenty-nine point nine seven million, you might not wanna call a spade a spade yet because two point four five of that are customers who didn’t explicitly cancel.
Their contracts just lapsed, and we wanna track that down here.” So this is something where somebody in account management needs to get all over that. And then obviously this is a demo data set, but it’s like nobody’s canceled. That would be a great problem for all of us to have in our businesses.
But you can see very quickly here how you can start getting much more granular, getting some new detail within here, having some really [00:59:00] to Chris’s point, you can start getting some rich pieces of feedback based on some of the open-ended context and things like that, that are possible with Claude.
Anthony, to your point, you were saying you’re using a term called skill. Do you label those skills so you can instruct the AI which one to use by name? There are a couple different ways to go about using skills. The best part about using a skill is that once you have a skill and you’re comfortable with the skill, like this is context that you can feed the data.
So it’s like think about it as a short-term or long-term memory. It calls that every time. You don’t have to explain it again. And so some people will build projects or they’ll build prompts where they say, “Hey, you know, I want you to do this repeated task, and when you run this task, I want you to call these five, these 10, these 15 skills to perform these specific actions every single time.”
So you call that like skill stacking. But it’s kind of like, let’s just say I do… I’m preparing my monthly close checklist, and as part of my monthly [01:00:00] close checklist, I want you to reconcile cash, and I have a skill for that. I want you to send a reminder email to these five people in account management because they have outstanding invoices and they’re on the AR aging.
That could be another skill. You could think about all the things that are on your checklist, and you could start building skills or automations behind each one of those. And this is where you start. Again, we’re shifting off the bike, boarding the plane, and starting to go much faster so that we can focus on our higher value work.
I hope this has been helpful for all of you. The stuff that we’re walking through here is, you know, it’s real. These are, that’s a real skill I use. One thing I would ask all of you to, on this webinar is shoot myself an email, shoot Chris an email. If you don’t have my email, shoot me a note on LinkedIn.
I’ll respond. Unless you’re trying to sell me something, ’cause we all get hit up with those all the time. But no. But seriously, we, myself, Chris, the whole, you know, we care that, we care, we want to help the broader [01:01:00] finance community learn this stuff. I think it’s new for all of us. But it’s also very exciting and there are some really powerful things that you can do here.
So this was, call it series one of our four-part series. We’re actually bringing on one of our customers next week who is doing some… You know, I just spoke with them this morning. They’re doing some really cool stuff with… They have a number of skills, and they’ve done some internal initiatives to encourage the use of these things.
So check in next week. Chris, thank you for joining this week and kicking off this series. And thank you all for joining us for the last hour.