How Credit Unions Are Using AI to Support Board Communication and Strategic Planning

Speaker A: So, Kirk Drake, welcome back to see you on the show. And Jeremy Wood, glad to have you here today. We are here for the first installment in our many episodes about AI in the boardroom. Uh, and Jeremy, you and the team at Innovations are one of the ones that I've watched for some time now develop AI use in the boardroom. And it's a really organic story that I think our listeners are going to really enjoy. And of course, our, uh, technology friend Kirk will, uh, help us to take it to many other levels. Uh, so in the first place, tell our listeners, uh, where you are today, what kind of work, uh, you are doing, and, uh, let's start there.

Speaker B: Yeah, sure. So, uh, I work at Innovations. I am the senior Vice President, Chief Strategy Officer here, uh, over enterprise planning, strategic planning, uh, growth, uh, I have responsibility for numerous support functions, and that would be it, marketing, hr, enterprise projects, and digital banking.

Speaker A: Wow, that's a lot of stuff. And apparently you're also the AI guy.

Speaker B: Yeah, I've been called that a time or two.

Speaker A: You leaving anything for the rest of them to do down there?

Speaker B: Uh, well, they're picking up. We're bringing some people on the team, uh, that are, uh, really excited to learn more about AI and, um, it's kind of catching fire error around here.

Speaker A: Yeah, that's. That's my impression, Kirk. Uh, go ahead.

Speaker C: Uh, so I got, like, 82 jobs. So, um, at its core. So, uh, Kirk drake crediting, uh, 2.0. Uh, and, and, and that's what we're talking about today. So we. We've, uh. I think my. I first saw an AI speech in 2015, which I think is like, uh, that really got me starting to think about this. Um, my AI book came out in 2020. Uh, I did talk about ChatGPT, although I had to go back and reread the book to find it. So, uh, it didn't have much. It had like, a little blurb about it. It didn't think it was going to be anything interesting. Um, in fact, I think it was Elon Musk making fun of it or something, uh, in the book. And then, uh, probably about three, four years ago, when ChatGPT first came out, really started transforming CU2. Uh, so, uh, that was a pretty big shift. That went from about 10 people on the team. Two or three are working on the core business today. The other six or seven are building AI platforms, tools, applications, all sorts of different things. You know, content, engines, you name it. And then we've spun up a AI coaching practice where we're Working with probably about a dozen different credit unions. I think our largest one is 7 billion or smallest one is 500 million. And we're uh, building those skills, uh, that Jeremy's talking about in their management teams and then uh, giving them the tools and platforms to go a lot faster and figure out, you know, what greenfield things we need to go after and what dinosaur eggs do we need to go kill. Uh, and how do we, you know, all kind of all around a long term trend of if someone with it's not AI is not going to kill you, it's someone using AI, right, Is going to take you out. And it's my belief 10 years it wouldn't be impossible to have a $500 million credit union with three employees. Right. M. And that's kind of a. And so the net interest margins are going to shrink in that and we've got to figure out how to become a lot smarter and how we do that and how do we move our transaction processors to member evangelists and cheerleaders and coaches and give a much higher touch service on that side, which is a huge transformation. Um, and so, and then in that work as well, a handful of those credit unions we've worked in the board room, both attending board meetings, pretending uh, to be a board member, using AI as well as helping them with the governance pieces, uh, doing educational sessions probably for about 50 credit unions on where it's going, what it's doing, you know, all that kind of stuff. And so, um, uh, kind of have examples from all walks of that.

Speaker A: We could have a podcast just on what you're, what you're trying.

Speaker C: Uh, we could take an hour.

Speaker A: And my favorite part, of course we've had many guests, none have attended from a food court. Let the first of all time groundbreaking Kirk Drake from the Detroit, uh, airport food court. Is that. Yes. Very nice, very nice. Well, uh, thank you for uh, adhering to our schedule because we're looking to get this content out. So uh, Jeremy, let's go back to the AI journey and innovations. Uh, and I know you kind of wrote up some timeline for us to go through. Let's just kind of loosely talk about the like, where did it originate from? Like, did, did this come from your personal interest or from strategy sessions? And then kind of how did you get, uh, started and beginning to apply AI? And which one did you use? Like just walk us through the story.

Speaker B: Yeah, sure. So I think that would take us all the way back until uh, late 2022, I believe. I think that's when Chat GPT actually came out. I, uh, started reading some articles about it, just trying to get educated. Um, but it wasn't until we attended the Governmental Affairs Conference in March of 23. Um, Scott and I actually sat in a breakout session. We were listening to a futurist talk about the future of AI and he was actually relating it to the medical industry, um, and how the AI is going to be able to synthesize large volumes of data, whether that be, uh, blood work or, uh, diagnostic test or imaging scans, and take that information and synthesize it and actually diagnose a patient. And, and um, you know, we were kind of fascinated by it. And I remember very clearly Scott looked at me and he said, this is going to be a disruptive technology in our industry, uh, one day, and we need to get in front of that and learn as much as we can. And so that's really kind of where it started. We came back to the office, we created an account, and then we just started trying to leverage the technology, uh, as much as we could at the time. Now, it was quite a bit different back then. Uh, it didn't have a web connection. You know, the knowledge base was a little bit stale. And then you hit some usage limits. So, um, you know, we, we've grown over time and as the technology's grown, we've tried to, to keep up with all of the updates and learn as much as we can about it.

Speaker A: Yeah. And you didn't give up. Right? You, in, in the, uh, in the beginning you had some difficulties. You talk, talk to me a little bit about what those difficulties were and kind of how you, you overcame them.

Speaker B: Yeah, well, sure. Even before the difficulties, I think we started out as users, uh, similar to probably most everybody else. I mean, it was kind of an initial replacement for Google. We would go and we would ask it questions and analyze the output. And we found pretty quickly that, you know, it follows the garbage in, garbage out philosophy.

Speaker A: Right.

Speaker B: Uh, a weak prompt, um, typically results in a weak output. And so over time, the challenges that we found is that we had to get really strategic and get better at prompting the model and providing the right context. And the better that we got at that, the better output that we, uh, received from the model and it made us better.

Speaker A: Okay, that, that. I think that is a huge takeaway. Kirk, you want to go, uh, any further with that? I think that is a very, very significant takeaway. I want to make sure our listeners hear that the quality of your prompt is good. Prompt is going to directly significantly impact the quality of your output. What do you have to add to that? Kirk?

Speaker C: Yeah, I think that's phase one, right? Like, I think you gotta go from how do I build, uh, really good prompts, right? And how do I get good at asking the right questions? What's my context, what do I want it to answer as, what skills do I need to have? You know, those kind of things. And then from there, uh, you got, you start getting into building GPTs, um, where you kind of have a repeatable set of prompt type things and then you got a little more training. Uh, give it some examples of what output you want, give it some um, uh, examples of inputs, ask specific questions that those become more reusable in the organizations. Um, so usually when our coaching will come m to the calls with 30 or 40 existing GPTs, and then we help everybody learn how to change and modify those for the specific credit union and then they've got things usable right away that they walk off into their business and they know how, know how to kind of get that. Tier 2 and then there's a tier 3 that we see beyond that where you start using it to actually build product, not just research things, not just do that, where you actually have like Claude code writing code, doing that straight

Speaker A: to recursive, of course. Kirk Break.

Speaker C: Exactly. And then, and then the fourth phase is how do I reimagine existing products being 100% AI centered in the first place, which breaks your brain because you start thinking about things in this linear fashion and it starts being a much more fluid dynamic. What's this interaction going to be like with the member, with the back office? What kind of analytics does it need to do? How do I measure quality? How do I give this insight? How do I take four chunks of disparate data and make it give you a version of truth that's more accurate than what you get in any one of them individually? And you have all the garbage core systems and other things that you're making sense of those, those in that, in that whole methodology.

Speaker A: So that, and that's where we'll end up. And we're, we're going to kind of, we're going to do a little more crawling before we start with the, the Kirk Drake Sprint. Right. So, uh, first, I love what you said, Jeremy, about treating it like a collaborator and pay attention to the prompts, not a more powerful search engine. And then talk to me about intentional use, like policies. You know, how did you, can you talk about like the next level that you took it to?

Speaker B: Sure, yeah. You know, uh, the technology is really good for helping you draft, uh, narrative type documents. And what we found is it wasn't a replacement for human thought. We had to understand the topic, we had to understand what we needed. Uh, we picked up the efficiency by not spending the time consuming portion of drafting all of the narrative content. Right. So we use, we leverage ChatGPT as a tool to help us build out the narrative based on the framework that we established internally.

Speaker A: So can you tell me more about the framework? Like uh, what, where, like anything about that, where did that framework come from? What was the subject matter? Uh, anything to help our listeners understand where, assuming that maybe some folks are like, yeah, we need to be doing this the very first steps.

Speaker B: You know, I guess when I say framework, really what I'm, what I'm talking about is, you know, you can't go to AI and just basically say, write me a security policy. You have to understand what needs to be in that policy. You need to understand the regulatory considerations. So what we would do is we'd build maybe an outline of what we were looking for, right? The topics that we wanted to cover, uh, the key players involved, uh, the regulatory expectations of that policy. We would build that into the context of the prompt and then use that to help us draft the actual narrative of the prompt. So that we weren't spending all of our time wordsmithing a policy, but we were focusing more on the elements of the policy and the important, uh, uh, concepts that needed to be in the policy without wasting too much time on the actual drafting itself, if that makes sense.

Speaker A: Mhm. Yeah. So go ahead, Kirk.

Speaker C: Uh, I'll add one thing in there that I find works really well and that I think absolutely the framework and being able to say, hey, I'm a credit union, I'm about this size, I'm regulated by ffiec, NCUA and this state. And by the way, I send out notices, uh, on text messages, so I probably have some privacy stuff I might have to do some. There's like 19 regulatory overlays that the credit union has to deal with in various, depending on the checking account or savings account or loan. Like it's insane. Um, so you give it all that kind of context, uh, and then, you know, say here's what I'm trying to do and here's who I'm trying to do it for and the quality of the response is going to get a lot higher. Now that's the super smart way if you actually know what you want way if you're lazy like me. I do. I Need to make a policy about something, ask me 10 questions to help me figure out what it is. And they don't come back with 10 questions that create the prompt that then goes and does that. And so that's the, that's the lazy ass way to do it.

Speaker A: All right, I want to say that again because again, I think that's a huge takeaway is when you don't know what to do with AI. Tell it who you are, what you're trying to come up with, and then ask it to ask you enough questions so we can help you to create it. You don't have to create it. You just got to give it enough context and it'll create it for you.

Speaker B: Yeah, ah, I completely agree with that. And a lot of times that's what we'll start with. We'll have IT interview us to learn more about us before we actually start, uh, providing the uh, the direction that we want it to go. So I think the more it knows about you, the better output that you typically get on the back end.

Speaker A: Now, uh, Jeremy, in your uh, list of things that you talked about, you got quite, uh, a lot of use cases. You mentioned, uh, policy drafting, which I also find it to be really good with the written word. Uh, strategic planning, support board level communications is what I really want to kind of unpack in this uh, session. Job descriptions and hr, uh, content. Anything you want to talk about in that area before. I really want to talk more about your special, uh, your special prompting.

Speaker B: Um, well, sure. I think the uh, the, the one area that we picked up some efficiencies in is drafting board meeting minutes. Right. It seems like a simple exercise, but where we started is we had somebody in the room that would record audio of the board meeting. Right. And so, uh, sometimes these meetings would last an hour, an hour and a half. Well, that person would then have to go back and they would have to spend a day, a day and a half going through that audio PA it, making notes, unpausing it, listening to more, and then using, uh, all of that, all those notes to actually draft the minutes of the meeting. Um, you know, the next step in our AI journey, we actually implemented an AI note taker to record the meetings and generate transcripts. And what we found is not only did it make us a little more efficient, but it allowed everybody in the room to pay attention to what was being said and being engaged in the conversation as opposed to trying to listen while you're also taking notes. So, you know, I'll, I'll get into that More when we talk about uh, where we kind of took that from a master prompt perspective. But that was one of our early wins where we picked up significant efficiency and then we've kind of leveraged that in some other areas as well.

Speaker A: Yeah, I want, I want to go further down that path. Kirk, you want to jump uh, in on the note taker because it seems to me like there's a, first you capture it and then there's a whole bunch of stuff you can do with it. What do you got?

Speaker C: Yeah, yeah, yeah. So, so my latest strategy, I have a note taker and video recorder that goes to every meeting. I uh, call it Kirkbot. It records all of those, all of that content come back, comes back and then it is used either in a sales perspective of like helping me draft the proposal to the person, helping me understand the board meeting, give me more context, what are my to do's, what are my follow ups, that kind of stuff. And then the secondary use case, all that video content and transcript goes through another tool that I built that takes all that video. It would zoom in on just me on this call. It will listen to the transcript for anytime I say something smart, which is not very often, um, but it will capture that, snip it down to a 30 second 90 second sound bite, create the four video formats for Instagram, LinkedIn, TikTok, etc. And then produce my prompt to it is produce it like I'm a middle aged man that's not very good at cap cut so it looks authentic. Uh, and then uh, puts the words on the, on the thing and then produces that outbound video content that I can just go paste on Instagram and say look, I said something smart and I didn't do anything different about my normal day. I'm just having conversations with you guys. It's creating those snippets and doing that whole piece. Um, so that's the, that's the full extreme side of it. Um, we use it in our fintech credit union kind of uh, advisory work. We record the calls, take the transcripts, um, read that and create the follow up actions and um, training insights for the fintech about why no one wants their product or why they're missing the mark or what's wrong with it. And then for the credit union executive we have a dashboard that they get out of it the exact same transcript which is here's uh, here's the things you might have missed, here's the, the where they actually are in their product development lifecycle M those kind of Things. And so they get real insights out of the exact same transcript, just using different prompts to figure out what actually happened in that meeting. And then my favorite hack, if you're good, is when you're having your teams calls and you're teaching people like something like, here's how to do a loan share transfer or something like that. You record those transcripts, use that to build your SOPs. Just upload the transcript and say, I need an SOP based on what I just taught. Unbelievable how effective that that SOP comes out first try, uh, almost no editing required, uh, out of that. And so the whole process of how do I document, produce, teach people how to do things becomes kind of second nature.

Speaker A: Wow. All right. I want to restate that one because I think that's another significant idea. Essentially what I think you're saying is, uh, I always tell, uh, our team to, if you're teaching somebody something, record a video of teaching someone something so we can use that video to teach the next person. This is the next level of that use. When you have your best trainer teaching something, take that recording, send it to your AI and turn it into, uh, all the things that you need to know, how you do things. All the, obviously the standard operating procedures. Right, that's what you're saying.

Speaker C: Exactly.

Speaker A: Great, great idea.

Speaker B: I just want to throw this out there real quick. Um, I'd like to expand a little bit more too on, uh, on what we do when we get that transcript out of our note taker. That's really kind of the catalyst, uh, for us developing master prompts and then master prompts turning into custom GPTs. Now, I don't know a lot about custom GPTs. We're still kind of novice users, so I think that would be a great place for Kirk to jump in and really provide some insight into the future of that.

Speaker A: Yeah, I want to make this for the per. For the board, the executives. They're just barely getting started. And I'm going to do many episodes on this subject. And the custom GPTs may, we'll tease it today, but I think that probably deserves its own episode. The thing I wanted to ask is, in your opinion, that when a credit union is just getting started, isn't the note taker maybe the low hanging fruit the easiest place to start to build your knowledge bank? Does that make sense or is there a. Is there a better place you think to start?

Speaker B: I completely agree with that. I think that was one of the biggest pickups, uh, we found in terms of efficiency. And it was super simple to implement, uh, you know, we registered, we logged in, we created the account, we set up the permissions and uh, it started joining meetings, uh, for all of us. And um, that's one of the things I don't know that I could go back and live without now.

Speaker C: Yeah, I never want to take board minutes again. Um, yeah, no, it's, it's, it's phenomenal at that piece of it. I think that's a great first step. I also think um, an even another great one for a board member is, hey, I'm going to this board meeting. What are three questions I should be asking, uh, of uh, that align with our strategy plan, um, that, that you know, I want to know a little bit more about. So be, I always say just be curious. Like that little thought that goes through your head is like, I wish I knew this. Just go check, type it in, chat it. I will say the downside of live use ChatGPT or any of these tools in the board meeting. It enables a board member to go to a micro level of understanding that is probably not appropriate for a board because they can get down to the nuance of a net worth ratio and what the regulatory framework is that isn't necessarily policy, isn't necessarily strategy, but could be weaponized in a board meeting to uh, you know, be like, well you said it was blah, blah, blah. And I think, you know, ChatGPT here says it's this and that becomes unproductive, uh, in a very quick way. Right. And so you got to find the right balance of how that's going to work. But it's certainly every board meeting I go into now I'm asking those questions of, you know, how does this align with this? How you know, what am I missing here? You know what, I didn't understand this thing and enables me to have a conversation about the board package before I even walk in the room.

Speaker A: Yeah, it seems like having that board package in uh, in the model and being able to ask questions of it is an interesting idea. And then the notes from previous meetings, uh, as building knowledge level of this board, the way it thinks, the way it doesn't think, the things that it's sees, the things that is missing. Like it seems like a lot of potential there. What comments do you have?

Speaker C: So one thing to be careful of. If you're the credit union, don't give the tool to the board to put the board package in to have the analyze it. That breaks your confidentiality and your attorney client privilege. Um, uh, really important point, they can do it on their own, with their own tools. But like, uh, and, and you know, you definitely want to have some board policy about what they can and can't do. There's um, some case law that isn't quite clear yet. Um, is, is if the board member isn't reading the full board package, are they doing their fiduciary responsibility versus having ChatGPT summarize the board package and do that? So there's definitely some nuance at the board level that makes it a little trickier and requires a little more thought, but certainly four minutes, ask me some questions like that kind of stuff. That's phenomenal.

Speaker B: I was just thinking, you know, in terms of our experience with the board, you know, some of the feedback that we received, we push a large volume of information to the board members, uh, you know, and, and what we found, uh, by using ChatGPT and AI is the fact that we can distill that information down into a, a meaningful level of executive summary that makes it easy for them to understand. And, and we'll, we'll ask Chad, uh, a lot of times, you know, read this, uh, from the perspective of a board member, you know, we use it kind of, you know, as a test case. Is this something that your average board member could understand? Do we need to expand more in one area? Do we need to condense it down? Is it too much information? While at the same time we'll also have it test, uh, from a regulatory perspective, you know, does this meet the expectations that the regulators would expect for us to deliver information to the board? And so we use it as, you know, as really a sounding board, um, a consultant to review the output of our work and make sure that it's appropriate and it's hitting all the key points. You know, I know regulators care about, uh, making sure the board is being brought up to speed on, you know, the camel characteristics. And so we make sure that we hit on those key points in our financial summaries and uh, and a lot of the documentation and we've, we've received good feedback back from the board that hey, this is much easier to understand. We appreciate the executive summaries, uh, and we still provide them the same level of backup documentation, but I think that they provide that higher level summary that helps them really digest a large volume of information in a short period of time.

Speaker A: Now have you seen it get, you know, it will lie elegantly, very elegantly. Have you seen cases, have you seen any cases of that? And uh, if you have, where did it show up and what did you do about it?

Speaker B: Yeah, we, we saw it early, uh, on we noticed that it would kind of fill in the gaps and, and make things up. Uh, we saw that in some of the summaries of the board meeting transcripts where it would rely on past conversations from past, uh, board meeting minutes. So what, what we ended up doing is we started working on developing master prompts and that was to help with, with formatting and tone and structure section headings, um, you know, consistency in the output. But we would also make sure in that prompt that we would uh, reinforce the fact that you cannot rely on any historical information. You know, the minutes had to be based on what was actually said in the meeting.

Speaker A: Right, I gotta stop you for a second. Tell our listeners what's a master prompt?

Speaker B: Well, I think it's something that we just kind of made up here at Innovations and, and it was, it was based out of necessity. Right. So as we would take that transcript from a board meeting, I, um, would feed it in and what I would find is month after month, the, the formatting would change. Now it's learned a lot about me over time and my writing style. But uh, from month to month, you know, it may reorder the headings, it may change the naming conventions, it may use acronyms where we prefer not to. And so what I found is I would have to remind it month after month, hey, don't do this. Right? We talked about this last month. And so uh, what we kind of uh, landed on is let's develop an overarching master prompt, building in the guardrails and the expectations, uh, from front to back on what we expect the output to look like. And so sometimes these master prompts would actually turn into a seven or eight page document. It would get to that level of detail. And so then when we would, would take the transcript and combine it with the master prompt, we found that the output got more and more consistent, consistent from month to month. Um, that were great for a while until we started trying to hand off tasks to other people. So I could provide the same master prompt to somebody else in the institution and they may get slightly different output. And um, and so to create that consistency, we actually started moving towards custom GPTs. And Kirk's uh, far more familiar with that than I. He could probably expand more on what that actually is, but that's where we're at today. We've kind of transitioned from that master prompt exercise into a full blown custom GPT that, you know, is consistent across the organization and user.

Speaker A: All right, before you take this to the ultimate. Just let's get really clear on prompting as a step.

Speaker C: Uh, very, this is very normal progression, right? You're going from I need to, I need to get good at prompting to I start having a prompt that I want to use on a regular basis and it needs to be consistent with output. And then your GPT becomes uh, maybe three or four steps further. I'm going to have the same master prompt. I'm going to give it some examples of what the output. So prior board minutes that I want it to look like. I'm going to give it some, a uh, handful. I might give it a tone and voice sample for the credit union that's not specific to Jeremy but is more specific to the credit union. Uh, I might give it a specific thing around who the board is which could go as far as to have LinkedIn profiles from every single board members and see a history of them. And you can, you can basically add ah, I think about 25 artifacts in this GPT. Um, it's changing all the time so maybe it's more now and the more of that you put in there is your training model essentially. And then once you have that GPT, Jeremy can hand it to someone else in the credit union. They don't get to see what's behind the curtain. All they see is hey, you need to upload your transcript. They upload the transcript, hit go. It might ask them three to five questions if you, if you've built that into the GPT and then it spits out the answer and it's going to come out way more consistent um, in that whole model now.

Speaker A: And is that uh, innovations who like, is that you that took it from uh, from the prompts to the custom GPTs. That kind of sounds like that's kind of getting into the area where your technology officer might be doing it. Or is that uh, is that Jeremy of all things?

Speaker B: No, that was actually me and it came out of necessity. Right. I uh, was trying to help our cfo. We were trying to create some, some high level financial analysis of how we were performing against budget and then the overall uh, condition of the credit union. So I had developed some master prompts. I had worked with him, uh, he liked the output. So when I handed that off to him to actually do on his own, you know, I get a phone call that says hey, this doesn't look anything like yours. You know, it's much more detailed than what we had dialed in together. And so you know, this became a back and forth between the two of Us. What's going on? You're using the same prompt. I am. Why are you not getting the same output? And so then, you know, back to the research phase. You know, why is this happening? And then that's how I ran across, uh, the topic of custom GPTs.

Speaker A: Right?

Speaker B: So then I thought, well, this is kind of like hard coding a master prompt into the background. And to Kirk's point, I was able to provide sample output documents that it could reference and tone and voice, uh, references that the model could lean on. Uh, and so when we got that actually in place, when I built that, I had him run it, and then it was immediately better than what we had done, but just strictly a Word document with a master prompt. And so that was kind of our progression. And so we built, uh, two or three now for different tasks. And I suspect that we'll continue down that path, uh, especially as we bring more users onto our platform. Ultimately, I can't do it all. I have to hand this task off or hand these tasks off to other employees. And, uh, and this kind of makes it plug and play for us.

Speaker A: Uh, so. Fascinating. So the, the big takeaway is prompting alone doesn't do it. I don't understand why. Maybe you want to tell us, Kirk, why prompting from another individual inside the organization with the exact same words doesn't get you the same output. Why is that?

Speaker C: Yeah, uh, because it's learning based on your style, tone and voice of what you've asked it to do before. And so it's going to recognize what Jeremy wants a little differently than what you want versus what me. And so the GPT forces it to, uh, a more limited view of the world that says, here's the tone, voice, style architecture of this thing. Don't use everything else you know about me when you're writing, right? Because like my, you can set in chatgpt your personal style. Mine's, you know, disruption with a side of humor use, sarcasm use, you know, all the normal snarky things that Kirk does. And that comes out in my writing style. And it's. I, I built a, like, one of the first things I did, I took both books, uploaded them in, and said, pretend you're a PhD English professor that's teaching kids, uh, and students how to define the voice, style, tone of a writer. So they need to learn how to write like Dickens or, you know, Bronte, uh, or something like that. Um, do that on My Credit Union 2.0 and financial book, uh, and then it, and then do it at a PhD level because there's a whole bunch of things to use to find style. Bose Tone. The style of voice and tone are three of like 20. Um, most of which the rest of us don't know because we're not English professors. Right. Um, and so it took that sample built the prompt, uh, and the. In the whole overview of who Kirk is and how I write. And that became my profile within ChatGPT behind the scenes and then everything. It's good enough and I've refined it enough. My wife can't tell the difference between ChatGPT writing and me. Um, with one exception, I can't spell and my grammar is terrible. So the quality that comes out, other than that, the jokes, the, you know, the rest of it are, are spot on. And it's very, very hard to tell the difference.

Speaker A: Virtual Kirk Drake coming to, uh, the Internet right away, I imagine you could

Speaker C: do that exact same thing on the credit Unstone invoice. And that becomes a GPT that's used to write marketing content. You could do a version that's a president's report, so it comes out the same way every time. You can do a version on it on an FPA analysis. You can do a version on prepping for your auditors. Uh, you know, and so each one of those things kind of becomes a business function GPT that's. This is a great way to learn the capabilities and what AI can do and begin to open up that world. I will say, in my experience, after a period of time, you start evolving to the next level of these, and you use these GPTs less and less, but they're really important in the direction of how to kind of build this and figure it out.

Speaker A: And you listeners may be saying to yourselves, I don't know how to make a custom GPT. And the lesson that you want to learn is when you don't know how to use AI, ask AI, how do I make a custom GBT so that I eliminate repetitive setup. I embed it with the knowledge of our institution and it'll tell you how to do it, right, Jeremy?

Speaker B: That's right.

Speaker A: So, uh, let's talk about, uh, what you are doing with it in Innovations now. How are you using it for, uh, you know, the most? Just, just give us the daily use. Where is it in your systems? Uh, you know, uh, what have you, maybe anything you put it in that you took it back from. Let's just kind of get into a little bit more of those details.

Speaker B: Yeah, I think we're trying to leverage it in every aspect of the operation. You Know, we talked a little bit earlier about drafting job descriptions. Uh, we absolutely use it in marketing. Uh, we use it as a, as an analyst, as a strategist. Right, to help us become better thinkers. Um, really, we, we every day we come up with a new use case for it. Uh, we use it a lot to pressure test ideas, right? We, we feed information in and, and have it critique our work, right. Critique our thought process. Where are the gaps? Where, what are we missing? Um, we use it, uh, in the finance area, right? We, we can load trended, uh, historical financial data, income statement, balance sheet, uh, ratios, uh, as really a high level analyst to help us understand weaknesses. Right. Strengths. We could do a full SWOT analysis on the organization. We actually kind of started there before we, uh, before we built our strategic plan this year. And we said, well, look at our financials, find things that we may be overlooking. Um, and that kind of gave us the framework to start building out the strategic plan to shore up some of those areas that we may have been deficient in, but we didn't necessarily realize because they didn't show up in a standard industry ratio that everybody's looking at for month to month.

Speaker A: Kirk?

Speaker C: Yeah, no, total. Those are. Can you hear me?

Speaker A: Yeah, we're good.

Speaker C: Okay. Yeah, I think those are great examples. And, and each time, again, just being curious, how can it help you? I use it a ton in, um, strategic planning, uh, and figuring out total addressable market, figuring out a competitive analysis. Um, when I come up with a new product or new feature, the first thing I'm doing is, hey, I'm using my McKinsey prompt. Pretend you're a McKinsey analyst and you're looking at this product. What are the gaps in what I'm looking at? How do I do it? Maybe I'll throw in, you know, pretend you're Doug English. What do you think about it? You know, it's really remarkable, uh, in all of those ways to, to have it, um, challenge you and, and have a really interesting conversation so that you come out of that with version three, version four, really much more well thought out and contemplating a lot of things. Um, one of my other favorite ones is how do I measure this? Right. Um, because there's a lot of things we try to measure that we really struggle with. And the reality is someone out there has figured it out. And it's been phenomenal at both figuring out ways to measure, you know, member impact, ways to measure, you know, those things and come up with scoring systems and those pieces that are really well thought out and work on first try.

Speaker B: Right.

Speaker A: Wow. Uh, so again, when you don't know, ask AI, how do I measure that?

Speaker C: That's the mind shift you got to go from. This is cute. I can research, I can, I can Google some stuff or I can chat some stuff to. This is an AI first world. My skills like learning PowerPoint and learning Excel are no longer relevant. That was not. I know we thought these were going to be things that lived with us for the rest of our life. They are not. The knowledge is not all that relevant. Wisdom is relevant. Um, and knowing how to define the problem and know what you want and knowing how to see opportunities. Because chat isn't very good at seeing things that it hasn't seen before. Right. Um, it can only be trained on what humans have created so far. And so um, it changes the entire mindset that once you go to, oh, this, this becomes a force enabler and a way of thinking about problems. And I need to start first there and then get into the rest of it. Not, not only use it when you're stuck, right. And you start to see a huge shift in productivity output. You know, uh, um, synthesizing information. You know, there could not be a better time for the amount of information we all have to deal with to have a much better tool to help us deal with it. Right.

Speaker A: So I'm gonna, I'm gonna read from Jeremy's summary document and Jeremy, I'm gonna do your. What has not been effective and then you uh, tell us about what has been effective. So uh, Jeremy said that what hasn't been effective is one line prompts. You don't give it enough context about who you are, what you're looking for. You don't get a very good output. Uh, blind trust in the outputs because it is the best storyteller of falsehoods that you have ever seen. I was uh, teaching some um, of our financial planners, uh, some ah, calculations around present value and I had built a little test and I had them do the test and I, then I gave it to a chat and was wrong, but beautifully wrong. I mean just so elegantly done, uh, numbers. I've, I have found it to be pretty uh, dubious in the area with numbers. I understand Claude is a little more effective, uh, with numbers. Uh, and what you said in your document about what had been really effective. Maybe you can tell us a little bit more about it. You said to treat it like a junior analyst or a partner. Like it a lot like a person that's in a role, uh, to, to build A reusable system like the uh, custom GPTs that we just talked about. And I think we're going to ask Kirk about some other episodes we might want to do on some of these things. Uh, iterating it, asking it multiple times. Don't just get the answer and go with it, ask it, back it up, show me your source data, what if that is wrong, that kind of thing. You don't just take it like a Google search, you push back on it and look for more depth of field. Uh, either of you comments on that?

Speaker C: I mean my favorite question when it gives me something is how confident are you in this, in this answer? And it's pretty revealing how many times it'll be like I give it a 40%. Like it's truth, it's truthful. You're like really like, uh. And so sometimes when you get something back, just asking that follow up question is a great way. Um, it has gotten so much phenomenally better than it was two, three years ago. Uh, so I would say I don't spend a lot of time worried about too much hallucination at this point. It's, it's much more accurate. Um, and I do 100% of FPA analysis. Anything finance related is Claude. I don't even bother with ChatGPT. In fact I probably only use Chat GPT about 5 or 10% of the time. And then I'm using Claude chat or Claude code, uh, probably 80% of the time. Things like gamma for PowerPoints and that kind of stuff, you know. So you end up having a boutique set of five or ten different tools, um, that you're interacting with, uh, kind of in that way. Last four or five business plans I've created with Claude. The financial models are phenomenal. It comes out with the spec, comes out with a code, you know, outline. Like it's, especially with 4.7 it is. I mean we are at the point in time when, when, if you go in and try it and you see what it can do, it'll it you're like oh, uh, this isn't science, uh, fiction anymore.

Speaker A: It's truly incredible. Yeah, yeah. I think for some of our future episodes we might go into the other models and when you might go there. So Jeremy, we're going to bring it to bring this to a close. Wrap it up. So any final list, uh, ideas for our listeners, especially in the beginnings of this process. Uh, like again, if you could go back, start again, how would you start? What would you do differently? What would you do? The same. How would you get started.

Speaker B: Yeah, I think the key is talking to it like it's a human.

Speaker A: Right.

Speaker B: Having real conversations with it, pushing back when you disagree on the output, asking it to be super critical. Right. Don't just agree with me, don't uh, try to make me happy. If I'm wrong, tell me I'm wrong. Um, that iteration, those follow up conversations to me are critically important to getting the best output. Um, and so I think that, that early on I would have, have asked uh, it to be a little more critical of my work.

Speaker A: Right.

Speaker B: It's nice to read uh, the initial, uh, nice uh, words that it says. Right. Oh, you're thinking about this the right way. But then if you say, well, you know, if I'm a regulator looking at this, um, how would you feel about it? And then it will give you a completely different perspective. Right. Which then, you know, causes me to say I need to dive a little deeper into this. I need to maybe consider uh, a few things that I haven't considered yet. Uh, but if you don't ask it for that critical feedback, you typically don't get it.

Speaker A: Excellent, excellent guidance, Jeremy. Thank you. Kirk, over to you.

Speaker C: Yeah, well said. Uh, I think the um, smartest thing I did early in this journey was uh, I sat down for eight hours on a Saturday and I said, let me just be curious about anything and everything I could try on this and just brute force my way through ChatGPT day one, uh, and this was back in 2022 or whatever. Um, and then the second smartest thing I did was set up a regular team meeting with group people at CU2 where we met every two weeks and talked about all the things we were trying. And I said something mean, uh, like hey guys, uh, it's not AI that's going to take your job, it's someone with AI and if I have to make a decision five years from now, that of who's going to come and who's going to go in this company. It's going to people that were tried and leaned in on AI. Right. It's not people that have just sat in the corner and pretended like it's not happening. So everybody's going to come every two weeks with something they're trying. I don't care how stupid it is. You can fail miserably. There's no judgment. We'll try to get better each week and we'll tease each other and laugh about how epically it fails at times and that learning curve. Even now I've got a group about 25 entrepreneurs that meets, uh, an hour every two weeks. People are doing brilliant things and everybody brings them, showcases them, um, we tear them down. We. And you can't walk out of that meeting without feeling like the biggest idiot on the planet. Right. Um, and there is something very powerful in the learning process of a group of people learning together that are challenging each other to get better and better. And it just speeds up that cycle, uh, exponentially. And you got 25 people doing deep research, not one. Right. And 25 people are always going to be better at something than, than one person on their own.

Speaker A: That sounds, that sounds like the kind of thing credit union group of people would do. Well, well, listeners, this is your opportunity to take this great idea from Kirk Drake and make it your own. Maybe, uh, create a group within your region, within your league, within your size, uh, and think about doing exactly that. Circle back on your time timeline and make your use of AI better and better and better. Jeremy, is the, the, uh, the, the prompt, uh, any of this content that we can put in the show notes, is that something that is appropriate to be able to share with the credit union community community, or is that a innovations thing that we need to keep at Innovations?

Speaker B: Well, I'm sure we could probably find a master prompt that doesn't create or doesn't contain any trade secrets that we could probably share with you, just to give you an outline of how it started for us.

Speaker A: All right, well, we'll put that.

Speaker B: What's been effective.

Speaker A: Thank you. We'll put that in the show notes. Most of our listeners are on YouTube, believe it or not, because they want to see that food court. Uh, and, uh, and Kirk, you are everywhere in everything. For the folks that are again on the beginning levels, what of your content might they plug into?

Speaker C: Yeah, certainly the AI policy is at a. That's the first starting point. Uh, and then probably, um, there's, I think we have a, uh, board governance policy and three or four things like that. And those are great starting points to kind of get into it. Uh, and then happy, anytime someone wants to, you want to get five or 10 people you're crediting together and do a free coaching session. I'm all in, happy to, you know, uh, have join that call, challenge you guys, push you, give you a bunch of ideas and uh, it's one of my favorite things to do. So.

Speaker A: All right, so Kirk, uh, the other thing I wanted to ask you to share with our listeners is about the other levels, because I've already got, got several other other episodes, uh, lined up and you're our uh, AI whisperer, if you will. Uh, so tell me, tease a bit about where we're going to take this in our future episodes.

Speaker C: Yeah, so I think, you know, I think we can do a whole episode on the, on the GPT generation and then I think there's another level behind beyond that of okay, we're starting to get some efficiency, People are a little less overworked, we're getting a little better. We've got our GPT that answers online banking questions or you know, whatever it is in that. Uh, so then I think it starts getting into uh, how do we start using it to solve legacy problems either by actually building a product end to end. So uh, uh, one example we built for Orsa credit unions DNF where a member, a non member, uploads a bank statement, it analyzes the bank statement, removes all the pii, figures out what fees they paid at the competing bank down the street, offers them a rebate, it tells them what fees they would pay at Orsa, and then offers them a, A, uh, uh, that they would rebate all the fees from the other bank to join the credit union that day. Um, and so built that product end to end using AI in about six weeks, had it live on the website. Right. So uh, you know, kind of start to imagine what's the world of possible about product innovation and how we can go target a particular demographic or use case in a very different way than we've done historically. And then the other side of that is how do we go find where we're using 5% of some SaaS application that we're paying a lot of money for, um, that we can do some serious expense savings by building some bespoke specific tools internally that never face members if we're not comfortable on the member facing side of it and do that side. And then I think the layer above that is, hey, that's great. How do I begin to build an ecosystem that allows the whole organization or a lot of people in the organization to be developing, building, launching, orchestrating these things and how do I keep that safe? How do I have microservices that begin to manage these apps and these tools and all that sort of stuff? How do I connect all the data, how do I make sure it's secure and all that kind of stuff so we can go to the full.

Speaker A: Sounds like a couple of episodes. I also have uh, one coming for you on governance, uh, with some highly, uh, some names that you're going to know, uh, in the movement to talk about governance because we want to do the this. We want to make the credit union movement successful at understanding AI and implementing AI, using it to, uh, to serve members better, push cost down, to increase efficiencies. We want to do it safe and in compliance. So we'll try to help you with that, uh, as well. Kirk Drake, thank you for coming back again. We'll see you again soon. Jeremy Wood, thank you so much for your innovation in AI use in the credit union movement. See you next time.