Ep 78 - Learn By Building: From Strategy Decks To Working Agents w/ Matt Bartles

Send us Fan Mail “We’ll learn AI once we understand it” sounds responsible, but it’s one of the fastest ways to fall behind. We sit down with Matt to argue for a different approach: learn AI by building with it, in small scopes, with real users, and with the humility to let the work teach you what the strategy can’t. The result is faster AI adoption, better judgment about what models can and cannot do, and a team that develops true operational muscle instead of slide-deck confidence. We dig ...

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:01Jason McGuinthe?

  Welcome back to Inspire AI, where we explore how artificial intelligence is transforming not just our technology, but how we think, learn, and lead. Today's conversation centers on a simple idea that's becoming increasingly true in an AI-driven world. You don't learn AI by studying it. You learn AI by building with it. And yet, there's a real tension here. Most organizations are still trying to understand AI before they use it. In doing so, they're falling behind the teams who are learning in real time through experimentation. So today, we're going to unpack what it actually means to learn through building, not as a hobby, but as a system,

• 0:46Jason McGuinthe?

  and ultimately, as a leadership strategy. Our guest today is someone who has spent their career not just talking about transformation, but driving it. Matt Bartles is an AI strategist with a track record of leading large-scale automation and AI initiatives across financial services, government, and consulting. He's built and scaled AI practices, stood up automation centers of excellence, and helped organizations move from manual fragmented processes to intelligent systems that deliver measurable results. Earlier in his career, he spent a decade at Capital One,

• 1:27Jason McGuinthe?

  where he led automation efforts that saved millions of hours of manual work and drove significant cost reductions. Since then, he's advised enterprise leaders, founded his own AI consultancy, and now operates at the forefront of enterprise AI adoption, helping organizations translate strategy into execution. What makes Matt especially relevant to this conversation is that his work sits at the intersection of experimentation and impact. He's seen firsthand that the organizations who move fastest with AI aren't the ones who study it the most.

• 2:10Jason McGuinthe?

  They're the ones who build, test, and learn in real time. So today, we'll explore what it really means to learn through building, not just as an individual skill, But as a capability leaders can design into their organizations. All right, Matt, welcome back to Inspire AI. How you doing today?

• 2:31Speaker 01

  Great, thank you for having me. I'm so glad to be back.

• 2:34Jason McGuinthe?

  Yeah, it's been over a year now since the beginning of this podcast journey of ours here at AI Ready RVA. I think you were probably one of the first five podcast episodes.

• 2:47Speaker 01

  I was really excited about the mission and to contribute. I love it.

• 2:52Jason McGuinthe?

  Yeah, yeah, we've come a long way in some ways, but a lot has definitely changed in AI since we last talked. So it's a fascinating place and I'm really excited to be talking with you about it today.

• 3:07Speaker 01

  The past couple of years have been some of the most fascinating, engaging and steepest learning curves. I can't stop thinking and talking about it.

• 3:15Jason McGuinthe?

  Yeah, yeah, I think last time I talked to you, you were working as a consultant. Am I remembering that correctly?

• 3:21Speaker 01

  I was working as a consulting at a beneficiary corporation that was called Impact Makers. And we, our biggest client was George Mason University.

• 3:31Jason McGuinthe?

  Oh, it was George Mason. Okay.

• 3:32Speaker 01

  Yeah, they were a great client, helping out with automation and starting AI several years ago.

• 3:38Jason McGuinthe?

  Okay, okay. And what are you doing these days? Tell us a little bit about yourself.

• 3:42Speaker 01

  Yeah, my, my journey into AI is weird, just like everyone's. I started out my career as a submarine officer, pivoted to financial services, was a business analyst at Capital One. Started out doing the machine learning models in 2010, which just look like toys today and moved through automation. After that, I pivoted to consulting, working mostly with automation. But then when ChatGPT hit in 2022, it was really clear that automation was going to be gobbled up by AI and eventually agentic AI. And wanted to harness that, I picked up some coursework and degrees and pivoted to AI strategy,

• 4:21Speaker 01

  where I am today at the Navy Federal Credit Union, helping accelerate AI adoption and create great use cases for our members.

• 4:29Jason McGuinthe?

  Nice. You you picked up some courseware degrees. You tell a little bit more about that. What's worked for you?

• 4:37Speaker 01

  I I have been a a lifelong learner. And so I'm always going back to the well to learn new things. And so at Capital One, we saw a huge well of technology and you've been a part of that. And so I went back and worked on a software engineering degree. And then as consulting.

• 4:56Jason McGuinthe?

  Like a real degree. Okay.

• 4:58Speaker 01

  Yeah, and a very generous education benefit to maximize that. And I spent two and a half years at the Harvard Extension School pursuing software engineering. And after two and a half years. Yeah, we kept balancing that with a full-time corporate job was great. But uh, the learning was exceptional, too.

• 5:15Jason McGuinthe?

  Yeah.

• 5:15Speaker 01

  After about after about two and a half years, I had enough credit to get a graduate certificate in a web application development. So I got that. And then really started to see my limits in in consulting. And picked up an MBA. I went to Darden for that. That was great. And Darden really started to embrace the AI revolution that we're going through now. And that was critical, a critical moment for me to not only understand where AI was going, but also understand how AI was working in the business context. So Darden did a great job of that. And now we're applying those things at Navy Federal, very exciting time. And every day, there's something new to learn.

• 5:59Jason McGuinthe?

  Yeah. Just out of curiosity, what was your concentration at Darden?

• 6:04Speaker 01

  Entrepreneurship. And even though I work at corporate job, some of those entrepreneurial lessons really are applicable in in the corporate context, too.

• 6:15Jason McGuinthe?

  Yeah.

• 6:16Speaker 01

  And say, like, one of the things they teach there is effective thinking. And how can I put together business ideas for the the assets I have under my control? And how do I be creative and create new value with that? That kind of thinking actually really works well in in corporate America, especially companies that embrace that kind of thinking. And in case I branch off on my own again, or decide to be in a corporate world, that that has really come in handy, especially with AI, because the tools are becoming more and more democratized. And so you can effectuate more things than you once could in a corporate role years and years ago.

• 6:55Jason McGuinthe?

  Yeah, well, the way I see it, even in a large enterprise, you're, if you're a leader, you're running a small business. And unless you're in a mature business model, running that small business in an enterprise, then you're building something, you're growing something. And so there's a continuum of growth and maturity that if you understand the life cycle of businesses, you're always on that, whether or not you're in a big company or a small business, or building something out of your garage, right? It is a true to form model that exists everywhere if you look for it. So.

• 7:31Speaker 01

  I, yeah. I love that. And having agency, like that, that entrepreneurial agency that I'm going to go do, this is in my control, I'm going to move it. That's greatly appreciated in a corporate context as well. So I love it.

• 7:46Jason McGuinthe?

  Indeed, yeah. Alright, so you've worked across enterprise, government, consulting. Where do you see organizations over indexing on planning and underinvesting on, let's say, building with AI?

• 8:03Speaker 01

  This is a really interesting one because we find ourselves making roadmaps for technology we don't fully understand. And let me be fair, the technology that we have today is capable of doing many things that we just haven't figured out yet. And so creating a three to five year roadmap for delivery of value in AI when it's drifting so quickly, that is a bit of over indexing. And so how have we pivoted to investing in building things with AI? It's really figuring out something that is feasible, deliverable in a short period of time, and can deliver value quickly.

• 8:43Speaker 01

  That's how we're building the subject matter expertise to work with these tools, understand what they can do, understand what value they can deliver in small scopes, beyond just what a proof of concept is doing. And that's helping enrich the learning cycle, how we can then do better planning. And so this is a place where the technology is moving so quickly, it's better to build and deliver value in small portions to figure out how can you create that better plan.

• 9:15Jason McGuinthe?

  Yeah. No, that really resonates with me.

• 9:19Jason McGuinthe?

  I've had a few guests on in the last six months and I get a chuckle and they do too. Every time I ask, where do you see the world going in three to five years? And they're like, that's impossible to even know right now because so much is changing so fast. You're just like, you just can't. So yeah, don't spend your time trying to predict where your business is going to be in three to five years or how to get there. You just have to, you have to be paying attention to the changes around you and incorporate them as fast as possible.

• 9:48Speaker 01

  To your, to your exact point, we're in a five-year planning cycle now for our enterprise strategy. And if you look back five years into May 2021, ChatGPT hadn't yet been brought to the mainstream. And it's sitting May 2025, there's no way I could have planned for what we're talking today. You have agents, working with tools and the power of the, some of the things we've been able to democratize, completely unpredictable and that ground has shifted so fast, there's hesitancy to plan that far out. There are some things that we do have to plan for, major changes into marketplaces.

• 10:33Speaker 01

  How do customers find and connect with products today is going to be very different in three years. And some of those relationships, we have to plan for now. So there is a place for long-term planning, but it is a murky place.

• 10:47Jason McGuinthe?

  Yeah. So what do you think the real cost is of waiting to fully understand AI before experimenting with it? Both at a team level and a leadership level?

• 10:59Speaker 01

  What is the cost of waiting to fully understand AI before experimenting is it's fully understanding AI is going to be a difficult task, given it's moving very quickly. Clouding that is we are still in a hype cycle for AI and agentic AI. There is many problems that need to be resolved. How do we govern it? How do we make it repeatable, transparent? What do customers want from AI experiences? All of that is still remains unknown today. If you wait to learn AI, when those things are figured out, at that point, you're so far behind the curve. It's going to be hard to catch up.

• 11:42Speaker 01

  And so by investing now, when we're still in the hype cycle, but these AI POCs are moving from science experiment to real, you will miss out on that learning cycle and be too far behind to catch up.

• 11:54Jason McGuinthe?

  Yeah.

• 11:55Speaker 01

  And so that is the true cost of waiting. And why we're investing and training and upskilling our workforce today, even though we recognize that we still remain in a hype cycle, is that cost of learning down the road is going to be too costly.

• 12:10Jason McGuinthe?

  Yeah. Yeah. And I've heard the opposition to that mindset as well, is where in a year or two, everything is going to be abstracted away and we're just going to be voice dictating everything that shows up on our screen and AI is just going to do it for us. So why do we have to learn about how it works right now? And I think about some people, if you're not, if you don't have the ability to leverage the AI systems as they currently are to experiment, I get it. Maybe you are, your company has to wait or your ideologies are just looking for excuses. But ultimately, knowing how these things work will benefit you in the long run, no matter what, right?

• 12:55Jason McGuinthe?

  I think about my experience in technology, the 20 years amassed, physical data centers, networking devices, really using switches and firewalls at my fingertips as opposed to abstracting it away in software. That was where I came from and knowing how those ones and zeros pass through the systems, that's, that is gold for even today's age because just because you can sit down in front of Amazon cloud and build a data center yourself, doesn't mean you don't need to understand what you're doing when you do it. You know what I mean?

• 13:32Speaker 01

  I agree. One practical example is we need to learn to manage AI costs and AI operations. And take for example, if you're running very heavy AI ops, that has a large token cost. And you had just mentioned.

• 13:54Speaker 01

  The cost of learning or the cost of delaying learning?

• 13:58Jason McGuinthe?

  Yeah.

• 13:58Speaker 01

  And where I was going was.

• 13:59Jason McGuinthe?

  The cost of waiting.

• 14:00Speaker 01

  Yeah, yeah. I had a practical example to add. And so as you think about the cost of learning, there's an actual monetary cost. And learning to manage AI op, what model are you using for a particular process? How many turns does it have? How are you loading context? What's going to be cached? What's going to be given to the model in each turn? How do you limit the amount of turns and what needs to be done in each turn? That system thinking can really reduce token costs. And if you don't go through that learning cycle, you are going to burn a lot of tokens unnecessarily.

• 14:42Speaker 01

  And that's unfortunately going to be the cost of your learning if you decide to delay.

• 14:47Jason McGuinthe?

  So you think the cost of leveraging AI is going to go up?

• 14:52Speaker 01

  Those lessons that you'll, those lessons will be waiting for you in the future. And learning them now is going to save you future costs.

• 15:00Jason McGuinthe?

  I see.

• 15:02Speaker 01

  But to your point, the cost of tokens is going to collapse. Now that's what the process we've seen over and over again.

• 15:08Jason McGuinthe?

  Yeah, they're pretty expensive using like Claude anthropic model APIs. Like that's just really expensive. But I'm lucky to have a company that allows me to leverage as many as I want theoretically. And the way, the way I leverage the coding systems I use based on a subscription, not a pay for token model. At least building apps, which is a lot more economical when learning, right? Like you don't have to, you don't have to use an anthropic API to learn. Uh, at least not the way I did.

• 15:45Speaker 01

  Yeah, I, where I run into token consumption costs is when the AI is part of the application itself. And how do you measure, take for example, uh, customer interaction on the phone with a voice agent?

• 16:00Speaker 01

  And so paying for, do I want to pay for GPT five opus? Do I want to pay for Sonnet? Which of those gives me the customer servicing level that I meets the need? Which one is the most cost effective? Where's the latency? Which one is the best at following instructions? That's an art that has to be learned for you to be able to scale and use agentic AI in your operation.

• 16:25Jason McGuinthe?

  Yeah.

• 16:25Speaker 01

  And it'll have a cost and that learning cycle is not going to be something. It will get easier over time, but having it sooner rather than later will expand the amount of things that you can do with your AI skills.

• 16:37Jason McGuinthe?

  Mm hmm. Yeah, it sounds like a very much a machine learning engineer type role to learn how to leverage the right models for the business value that you're looking to abstract.

• 16:51Speaker 01

  It's something that if you work with consultants or some of the Microsoft or X AI, I suppose, Claude, I haven't worked with Anthropic. They want to push their model, but some of their models aren't great at intelligent document processing. Right? And where the Microsoft has, what do you call it? Intelligent document processing, which it seems to be the gold standard for IDP. Whereas if you were to use Claude or Chat GPT for that, it's not nearly as robust. And so model selection is something that is a bit of an art. And if you develop it as a corporation, then you're not reliant on consultants or some of the labs to help you out.

• 17:37Jason McGuinthe?

  Yeah. What about leveraging the models on Hugging Face, the open source marketplace for models? Can't you find one that's built specifically for intelligent document processing? If that's the case you're looking for as opposed to paying Microsoft?

• 17:57Speaker 01

  This is a great question.

• 17:59Jason McGuinthe?

  You have to be able to host it yourself.

• 18:01Speaker 01

  Yeah, you got to host it and it depends on where your business is. And so, for example, if you're a Microsoft top, you can use the, I think it's called the model garden.

• 18:11Jason McGuinthe?

  Where this is an Azure service.

• 18:13Speaker 01

  Yeah, you can take a model, essentially Azure has a copy of Hugging Face.

• 18:20Speaker 01

  That has a bit of a, what would we call it, curated selection of models? It's still huge. I think it's in the 10,000 range. So if you wanted to do text to image, image to text, intelligent document processing, that's all hosted on Azure, but the token cost remains. In fact, you might even pay more on the Microsoft than you might into AWS using a similar model. And so that whole model selection, how do you get it?

• 18:46Jason McGuinthe?

  That's the art. Yeah.

• 18:47Speaker 01

  It's going to be an art. I would say primary to that, though, is taking a process, say, underwriting, right? Banks underwrite millions of loans a day, if not more. And if you can reframe that in the agentic, in agentic operations, that's a hard thing to do, too. Real estate lending is enormously complex. You have a bunch of documents. They have regulatory requirements. You need to zip in two months of your banking statement, your, what is your 401k level? All those documents are different. And so picking the right model that can scan that, picking the right model that can harness that data

• 19:28Speaker 01

  and put it in an underwriting model, picking the right model to interact with the customer, that's, that's the art and it exists in the frame of what, how the business operates.

• 19:39Jason McGuinthe?

  Wow. Yeah.

• 19:40Speaker 01

  Yeah. Yeah.

• 19:41Jason McGuinthe?

  That's very insightful. It's definitely taking us down a rabbit hole.

• 19:46Speaker 01

  Oh, yeah.

• 19:47Jason McGuinthe?

  For where I was headed, but it's good, good knowledge to share for sure. And, but I need to bring us back so that we can stay focused on our theme for today, which is developing experience, right? And, and building a muscle of learning fast, right? So in your experience, what do teams actually learn from building that cannot, that they cannot learn from, let's say, strategy decks, courses or certifications?

• 20:19Speaker 01

  This is a great question and timely because I recently taught a class on agentic AI.

• 20:26Speaker 01

  In this class, we use a co-pilot agent builder, which is available in the M365 suite. And we all built a meal planner agent. So what the agent will do is go into a recipe repository, help you pick out recipes for how many family members you may have, and then build a meal plan with cooking instructions and a shopping list.

• 20:50Speaker 01

  And that was a relatable agent you could build for people who all have very different roles.

• 20:57Jason McGuinthe?

  And how do they access the agent? Is it a UI that they talk to the agent with? And they're like, build me a seven day meal plan. Is that what it is?

• 21:06Speaker 01

  Exactly. And so you would create agent instructions. And so you define its purpose. You are a meal planning agent. And the first thing instructions is collect, for example, if you have any food or dietary restrictions, like allergy, food allergies. Do you have food preferences? Are you vegetarian, for example? How many people are in your family? What meals do you cook? What's your meal preparation like tolerance? Can you, do you like spending eight hours making meal? Or you want to do quick meals with five ingredients? And so that was an interactive part of the AI. Then the AI will reach out and Python and, or I'm sorry, will reach out to the repository to figure out meals that possibly fit that.

• 21:51Speaker 01

  There's some negotiation with the user and then it'll create, use a Python to create a document that has a shopping list or whatever. So the question is, what do we learn at the team level? They start to learn, what does that agentic experience look like for them? How do you plan meals? How much time do you want to spend actually picking the meals? I don't really care about picking the meals. I don't have a, I'm not very selective, but some people really are. And they really want to have, spend more time iterating on the meal. And so through this process, they understand, hey, what does an agentic experience feel like? And then they also learn, what are the limitations of the agent?

• 22:32Speaker 01

  And here's an interesting one. So an agent, if you look at a recipe repository, you've got spicy chicken, crispy chicken, you know, chicken al fuego, like hundreds of different kinds of chicken recipes. And the AI picking from a long list of very similar things seems to fall apart.

• 22:49Jason McGuinthe?

  Does it?

• 22:51Speaker 01

  Yeah, if you, if you.

• 22:52Jason McGuinthe?

  Not able to predict exactly the right one, that sort of thing, and then comes out of left field with the, its recommendations?

• 23:01Speaker 01

  Exactly. And so there's 55,000 recipes on, in the recipe, the repository. And there's, of that 55, there's 100 different kinds of pasta that are very similar to an AI, which is like, hey, and so if you randomly pick up a pasta dish.

• 23:16Jason McGuinthe?

  Sure.

• 23:17Speaker 01

  And so that's, and that's a limit.

• 23:18Jason McGuinthe?

  Random so much as it's statistically chosen, right? Yeah.

• 23:23Speaker 01

  And three people could ask it for a same meal plan and they'll get three different kinds of pasta.

• 23:27Jason McGuinthe?

  Meal plans, yeah.

• 23:29Speaker 01

  So it develops some inconsistency. And the whole context of this is to understand the limits and, and, and building agentic experience. And so that teaches them i- in the banking context, hey, how many different pathways do I want to put the mem- the customer on? Because AI would select between 100 different pathways. Today, we have the over several hundred different work a member can go through. And AI is going to struggle, at least in its current iteration of technology, to pick the right one out of the list. Credit balance refund of overdraft fee, credit balance refund for a fee waiver, and like all those things. And so we have to redesign the workflows in order for the agentic AI to operate quickly.

• 24:12Speaker 01

  And so there's so much learning to be had when you have your hands on. As an individual, I don't, I can't learn all that on behalf of the team. And so by teaching teams agentic AI, they find those limits on themselves. If that makes sense, if that makes sense.

• 24:27Jason McGuinthe?

  Yeah, that makes perfect sense.

• 24:31Jason McGuinthe?

  Do you, do they ever consider how to put like rails on those systems that they're building where they're getting more deterministic with the model outputs versus just leaving it so open-ended that it's going to continue to fall apart no matter what?

• 24:48Speaker 01

  This is part of the art of it.

• 24:51Jason McGuinthe?

  Yeah.

• 24:51Speaker 01

  And it's the art and science of the AI itself is probabilistic. And the tools are deterministic and the data can be either or. And so how do you balance the probabilistic AI with deterministic things? And this happens in real work, especially in banking, right? And another example is an AI that pays credit card payments, for example. So you can't pay a credit card, for example, on a, from a savings account that has zero dollars. You'd overgraffit, right? And so if you don't have sufficient funds, you shouldn't pay that account. If you just put that in context of an AI, do not accept a payment from an account that

• 25:35Speaker 01

  has insufficient funds, it will follow that rule 90% of the time. It's a probabilistic model. And that's tested, you can go to IS then and test that. And it'll enforce that rule 90, 95% of the time, even the best model.

• 25:50Jason McGuinthe?

  Right.

• 25:51Speaker 01

  And so you need to have a deterministic tool in your payment API or payment MCP, for example, that enforces that rule. And so teams have to learn, what are they willing to accept in the probabilistic experience? And what do they have to accept in the deterministic experience? That is an art and a muscle that is still being built.

• 26:10Jason McGuinthe?

  Yeah, yeah. So you've scaled automation and AI across large organizations. How do you turn experimentation from something a few people do into a capability that the entire organization benefits from?

• 26:26Speaker 01

  I will not claim mastery here. This is something that scaling a capability is difficult because many of the business units don't specifically want to align to a single capability. And let me give you an example. Intelligent document processing. So many business units are using intelligent document processing. Collections, underwriting, legal, audit are all using intelligent document processing. Supplier management, intelligent document processing, right?

• 26:58Speaker 01

  And so, part of the art is creating a capability that it can both be configured to meet a specific need, but also serves that need very well. And I want to be very honest, as we scale these things, we find out that these reusable capabilities really only get people thirty percent ahead, but that thirty percent isn't nothing. It also lowers total cost of managing it, the total cost of ownership. And so, by creating these reusable capabilities and marketing them to the business units and showing success there, we're able to align them to capabilities. And that has a lot of value for our IT partners who are managing, instead of managing five

• 27:39Speaker 01

  or six IDP systems and disparate software, really have them managing one or two different configurations of a single capability.

• 27:49Jason McGuinthe?

  Wow.

• 27:49Speaker 01

  But there's a lot of human, there's a lot of human messiness in here. Some of the processes have to change, some of the mindset has to change, and so that's the, the part we have to play as strategists is to massage to get them to use a central capability versus buying, buying the fast partner.

• 28:07Jason McGuinthe?

  Wow.

• 28:08Speaker 01

  By the way, in terms of these capabilities, this is a really interesting one, as we look across complex organizations, they have so many duplicative ones. We have fifty different knowledge management systems, if not more. And so, knowledge management is another one where you can have a centralized capability that can be reconfigured and spread across the organization. But it is humbling, that is a very humbling work, because every use case is quite different, and alignment is hard.

• 28:36Jason McGuinthe?

  Yeah.

• 28:37Speaker 01

  So, it is a goal that is worth pursuit. It is an art.

• 28:42Jason McGuinthe?

  Yeah. Yeah, you almost have to build for your eighty percent of your customers and then worry about the twenty percent later.

• 28:50Speaker 01

  Yeah, one of the things I have discovered success here in automation, and I haven't replicated this in AI, is making sure the onboarding process is buttoned down. Hey, we'd like to onboard you onto our intelligent document processing capability and having processy documents and that muscle memory. So the customer of that process feels secure and heard and embraced that it's the soft part of getting somebody on a capability and then making it low, low cost of management they like. And so it's a soft skill. That has a very technical piece to it. It has to be made in a configurable way. That's not always easy.

• 29:29Jason McGuinthe?

  Yeah.

• 29:30Speaker 01

  And there has to be an honesty, a piece of honesty in there that 30% of it can be reused. You get 30% lift, but the business gets a big, the business and the overhead cost gets really reduced. So.

• 29:41Jason McGuinthe?

  Yeah. Yeah. Have you seen teams build a lot of things, but not really actually create value?

• 29:48Speaker 01

  Oh, very much, very much. And what is value? Some of these are just lessons learned. Some of these are lessons learned. And I'll give you an example recently.

• 29:57Jason McGuinthe?

  Fail fast, learn quick.

• 29:59Speaker 01

  Yeah. And the thing is, these agents, so that we have a democratized agentic program where 12,000 people can create agents. And so they're creating agents fairly quickly, but as quickly as they create them, a lot of them are going in crap. And so Darwinism is happening really quickly in agents. But what the pattern that's happening is people experiment, they don't produce a lot of value, they're deleted, but some do. And then they start to get shared. And we have dashboards to say, wow, some of these agents start to get really popular. And then we layer on additional.

• 30:33Jason McGuinthe?

  Oh, wow. Yeah, that's great.

• 30:35Speaker 01

  Yeah.

• 30:36Jason McGuinthe?

  So build whatever you want. And if it, if you find use out of it, we'll identify that usefulness, then we'll share it out and then it'll become part of the marketplace that we maintain.

• 30:48Speaker 01

  Exactly.

• 30:49Jason McGuinthe?

  Yeah.

• 30:49Speaker 01

  Exactly. And that process has to happen really quickly. And so there was a McKinsey report where their organization has 74,000 agents actively being used. And that doesn't appear to be well managed from a banking standpoint. How would we defend against that?

• 31:05Jason McGuinthe?

  No.

• 31:06Speaker 01

  So we, M365 does a great job of helping us with dashboards to take a look at that. And when we constantly monitor that. And it's interesting to see what agents are created, intelligent document processing, how they, that's a frequently built one. But really like two or three of them are actually really good and those are the ones that get shared and propagated across business units. And then we can take the best practices out of that one and then advise businesses to stop that one. And then when we get to a certain user base, additional governance gets layered on. And the reason for that is it becomes a higher risk because it's scaling back. As if, if you remember the Excel macro era where that was being, macro was the backend of a lot of business calculations, but it wasn't very transparent.

• 31:51Speaker 01

  We're not going to make that mistake again with our AI program.

• 31:54Jason McGuinthe?

  Yeah. Fascinating. To that end, how do you guard against teams thinking they understand something when they actually don't? Like building agents, oh, it's so easy. I just use this template and then I have this automation in front of me. But what then? Like it, it becomes garbage quicker than it gets put to use, right? So how do you guard against that?

• 32:19Speaker 01

  Ultimately, every associate is held accountable for their outputs, whether it was generated from AI or not. And the training goes with that. At the individual level, the agent creator, you always are liable for the outputs of that agent. And we trust our associates and in good faith, they have been throwing away the bad agents. Yeah, this didn't work and there's a bunch of, and I'll give you an example. The AI agents we have today don't work well with Excel like people think they do. People see Excel and they think of it like data and they're like, oh, I'll upload to an AI and AI will treat the Excel like data. What actually treats it like text. Does that make sense?

• 32:59Jason McGuinthe?

  Text, yeah.

• 33:00Speaker 01

  If you don't configure it right. And so a lot of people are thinking that, hey, it will work well with Excel. I can use it to manage tickets. And the AI we have just doesn't do that well. It just, it falls apart. And so they quickly self diagnose that and throw those Excel based agents out. And so there's some self management. And then on top of that, the business unit risk officers, each business unit has a risk officer or trained in AI and agentic risks. And on top of that, our manager and leader layers are trained to, hey, how did you use AI in this process? And that's, there are two sides of that coin. One of them is that creates the environment for people to share their AI games with their major, with their manager.

• 33:46Speaker 01

  But also it reinforces that you remain liable for the outputs of your experience. So

• 33:52Jason McGuinthe?

  Gotcha.

• 33:53Speaker 01

  Soft with the answer. We do have hard deterministic rules in place for how those agents can be used, like PII and PI data. And then we have trust and governance, of course.

• 34:04Jason McGuinthe?

  So at what point does a leader need to shift from personally experimenting to designing systems that enable others to experiment effectively?

• 34:15Speaker 01

  Oh, man. I recently read a book on, it was called AI Derived Leadership. How do we, how do you lead in the age of AI? And it's, it's, it's an interesting question.

• 34:30Jason McGuinthe?

  Do you recommend that book?

• 34:31Speaker 01

  If you were new to AI and you're just starting to your AI journey, it may provide value. And so to a new person, it may be an a book.

• 34:40Jason McGuinthe?

  I see, yeah.

• 34:40Speaker 01

  To me, it seems like some of these things we're already experiencing. If you're in an, in an organization that's a bit behind, I would recommend it. And so maybe I was a bit.

• 34:49Jason McGuinthe?

  Sorry, I didn't mean to get you off the, your thought track on

• 34:52Speaker 01

  Oh.

• 34:53Jason McGuinthe?

  On the question about experimenting effectively.

• 34:57Speaker 01

  Yeah, I think it's important for leaders to experiment as well and get comfortable with the tools so they can effectively challenge the decisions and creations that their teams are building.

• 35:09Speaker 01

  And so I don't think there is a point where experimentation has to stop. Though I understand, hey, it's onerous on a leader to have to get their hands dirty when they have many other things. And so potentially you get to a point where you're experimenting enough to ask great, insightful, deep questions. And if once you have that sense as a leader, then I think I can effectively challenge what my team is creating and their approach and how they're integrating AI. At that point, you're learning, you're at status quo. And that's where, that's a place that will, leadership will have to maintain. It's moving fairly quickly. And so whether that be experimentation, going to conferences, working with subject matter experts, leaders have more things at their disposal.

• 35:52Speaker 01

  But they should be able to effectively challenge their team versus being a pass through for risk.

• 36:00Jason McGuinthe?

  Got it. So Matt, my final question for you would be, if you walked into a company today as an AI strategist, what are the first systems or behaviors you'd put in place to accelerate learning through building?

• 36:18Speaker 01

  This is an exceptional question and a lot of companies probably find themselves asking this. How do, what are the first systems and behaviors we need to reinforce? I think of things in terms of platform, people and processes. And so the systems would have to be designed to do those. And so amongst the first system is getting AI in the hands of your teams as fast as possible in a well-managed, well-governed way. And so if there's a, in the Microsoft domain, you can pay for $20 a month, a Copilot premium license. That's a great first start to get your people starting to interact with AI.

• 36:58Speaker 01

  Then you need to start understanding your processes and where can processes leverage artificial intelligence. So that's the process piece. And the platform piece is really tricky now because many of the platforms are evolving fairly quickly. But getting your tech teams ready to work in a, with AI. And some of the things we've discovered here is the long pathway. Most software development was deterministic and had very easy testable verification. And AI is a bit more slippery. And starting to get the teams having an SDLC that accounts for that would be the next system I could have played. I'd love to hear your thoughts on this though, David.

• 37:42Jason McGuinthe?

  The three Ps as a framework didn't actually jump out at me until you just started explaining them. I think it's, it frameworks are brilliant, right? Keeps you grounded in what you should be doing. When I think of managing systems at a large scale, I think of the technology, the processes and the people, right? And so I think you probably just reframed those three things that I use all the time personally as your own. You got the platform, right? And a lot of companies are becoming platform-based companies these days, which unifies the businesses around specific centralized functionality. And that's important because you don't want

• 38:25Jason McGuinthe?

  10 different document management systems, right? You want one or two for, based on the use cases, and you want to be able to fold in all that content so you can easily access it at any time, right? So the platforms are super important. I think if technology is going in the way of platforms, and so that in my mind, it's all the same. Like, how does all this change the SDLC, if you will?

• 38:50Speaker 01

  Okay.

• 38:51Jason McGuinthe?

  So I'm a lifelong learner as well.

• 38:55Speaker 01

  Yeah.

• 38:56Jason McGuinthe?

  I've, I've spent many a months building projects for the sake of learning, or in my learning process. And I also build systems for nonprofits as well as enterprise scalable systems. And I think that this particular episode has been really interesting to me because I truly believe that spending 30 days to learn a skill is no longer the world that we live in. You learn the skill by doing, especially if you're in or around technology. So think about learning now as a new behavior. It's not about opening a book and reading about it.

• 39:41Jason McGuinthe?

  It's about opening an interface and trying it. Right? And that's how you learn, right? You're going to learn by experimenting. That's why I like this episode so much is because I've realized that over the last six months, that you Decks of the world, Coursera, deep learning AI, like all of that serves a good purpose. And I can go in, I can learn really small snippets, but I'm not going to sit there and finish an entire course because I, one, I don't care about the certificate at the end. But two, I'm going to pick up what I need from a few videos and then I'm going to go and build the thing because the systems that I'm leveraging now allow me to do that in such an accelerated pace.

• 40:26Jason McGuinthe?

  And while I'm experimenting, I don't have to worry about the evaluations of these systems. I'm trying things to see what works, right? And one of my friends, we get together weekly on, on building systems for our nonprofit AI ready RVA. And his take was that we're building so much so fast that that's our learning process, but it also allows us to figure out what works and what doesn't, much like your class that you shared with your associates or whatever, teaching them how to build an agent or use an agent by building one themselves and teaching them the limitations of that's exactly it. Right? You have to get your hands on that stuff. And, and that's where we live in right now.

• 41:10Jason McGuinthe?

  So it started with software because software is the closest to ones and zeros where these agentic systems know about the world around them. They don't know the language unless we teach them the language. And through their language, we taught them our language. And now they're really good at our language and they're really good at coding and accessing the systems because of how we taught them. But it's going to grow. It's going to grow into so many more fields and those fields are going to be impacted just as much as software engineering has been impacted and data science and in the field of machine learning. So everybody's going to be learning at this pace. They just may not be realizing it yet. So there you go.

• 41:53Speaker 01

  I'm excited about it. Jason, thank you so much for having me. I really appreciate your time and I'm in love with the AI ready RVA. Always happy to come.

• 42:03Jason McGuinthe?

  Awesome, and I will have you back again, I promise that, so.

• 42:07Speaker 01

  I think the world's gonna be, wait another year, who knows?

• 42:10Jason McGuinthe?

  Yeah, it'll be an exciting. Catch up, for sure. Thanks, man.

• 42:15Speaker 01

  Nice talking to you.

• 42:32Jason McGuinthe?

  This episode of Inspire AI Podcast is sponsored by Exponent 21. I keep hearing the same thing from business leaders lately. We're doing the work. But leads are harder to explain. Traffic is down, attribution feels fuzzy, and the old answers don't really fit anymore. What's changed is how people search. Instead of clicking links, people are asking questions, and AI is answering them. Tools like ChatGPT, Google's AI-powered results, and Perplexity are shaping which brands get mentioned, referenced, and remembered.

• 43:15Jason McGuinthe?

  Exponent 21 specializes in AI SEO. They help businesses understand how to show up consistently inside those AI-generated answers across questions about their industry, their services, and their brand.

• 43:33Jason McGuinthe?

  What stands out to me is that when people ask AI tools, "Who's leading in this space?" Exponent 21 and its leaders are often referenced near the top. That comes from doing the work and shaping how this new search environment actually functions. To understand what this shift means for your business, visit exponent21.com.