Skip to content
When Agents Have Wallets, Trust Is Currency episode artwork

May 21, 2026 51 min

When Agents Have Wallets, Trust Is Currency

from Invisible Machines podcast by UX Magazine Hosted by Invisible Machines

Mastercard's central AI team receives roughly a thousand requests a year from across the organization. A few years ago, most of them were for chatbots. Today, most are for AI agents. Federico Cohen Freue, Executive Vice President of AI & Data Operations at Mastercard, has watched this shift in real time and knows exactly what it reveals about how enterprises are (and aren't) thinking about AI. In this episode, Federico explains why the name people use for what they want matters less than whether they understand the conditions that make it work. “Ball bearings,” as Robb Wilson puts it: demos can't reveal the difference between a solution that will hold and one that will blow up the engine. What actually matters is training, fluency, and a clear framework for where to deploy AI with purpose. For Mastercard, that framework is deliberate: use AI to make commerce more secure, smarter, more personal, and to make the company itself stronger. Not everything. Those things. The simplicity is a feature, it gives a sprawling global organization a shared language for prioritization and a stable center as the technology keeps evolving. In the second half of the episode, Robb and Josh share a demo of an AI-first approach to knowledge management and learning. Rather than asking people to query a knowledge base, the system proactively teaches, building a knowledge twin of what someone knows, identifying gaps, and using a traveling salesman approach to map personalized, dynamic learning paths. Think GPS for expertise: here's where you are, here's where you need to go, turn by turn. Federico's reaction gets at why this matters beyond the demo: it's not a technology question, it's a cultural one. Teaching people to engage with knowledge differently is the harder transformation. And it's the one most enterprises skip. The discussion makes it clear that trust, knowledge, and agents that know what they're doing before they're sent out to do it are the throughline. ---------- Support our show by supporting our sponsors! This episode is supported by OneReach.ai Forged over a decade of R&D and proven in 10,000+ deployments, OneReach.ai’s GSX is the first complete AI agent runtime environment (circa 2019) — a hardened AI agent architecture for enterprise control and scale. Backed by UC Berkeley, recognized by Gartner, and trusted across highly regulated industries, including healthcare, finance, government and telecommunications. A complete system for accelerating AI adoption — design, train, test, deploy, monitor, and orchestrate neurosymbolic applications (agents). Use any AI models Build and deploy intelligent agents fast Create guardrails for organizational alignment Enterprise-grade security and governance Get in Touch: https://onereach.ai/contact-us/?utm_source=soundcloud&utm_medium=social&utm_campaign=podcast_s7e10&utm_content=1 ---------- The revised and significantly updated second edition of our bestselling book about succeeding with AI agents, Age of Invisible Machines, is available everywhere: Amazon — https://bit.ly/4hwX0a5 #InvisibleMachines #Podcast #TechPodcast #AIPodcast #AI #EnterpriseAI #Mastercard #AgenticAI #KnowledgeManagement #AILearning #AIStrategy #AIAdoption

Episode details come from the podcast's RSS feed.

Embed this episode

Ready to play

When Agents Have Wallets, Trust Is Currency

0:00 51:16
of matches

Transcript Auto-generated, may contain errors

Welcome back to Invisible Machines. I'm Josh Tyson, joined as always by Rob Wilson, co-founder and CEO of OneReach.ai. Rob and I are co-authors of Age of Invisible Machines, the first best-selling book about agentic AI, and this is the podcast where we explore how intelligence is actually built and where it's going through conversations with the people shaping it. What changes when an organization or a consumer hands an AI agent real buying power? According to Federico Cohen-Frua, executive vice president of AI and data at MasterCard, when agents have wallets, trust becomes currency. What is this agent designed to do and not to do? How will every party in a transaction see what's happening? These questions get harder to answer without a shared understanding of how AI actually works. Fed's central AI team fields roughly a thousand requests a year, but an ask for a personal assistant can reveal gaps in how people view the technology, which is why Fed is focused on fluency before deployment.

To that end, we show something aimed at that second pain point, an AI agent designed to educate users about AI by mapping canonical knowledge, assessing what they already know, and building a personalized path forward. Watch this episode on our YouTube channel to see this teaching agent in action and enjoy this conversation with Fed Cohen-Frua. Fed, so last time we talked as we were preparing for this episode, you were mentioning that in your role at MasterCard, you sometimes get like, I think you said like a thousand requests every year for different applications of AI that people across the organization. That is like an insane fire hose to drink from.

It is, but there's so much demand for using AI across the entire enterprise. To some degree, I think I'm almost glad that we have a lot versus not. Let me give you some background in what it means. I work in our central AI and data team. It's our hub. It's where a lot of our very dedicated talent on the topic is focused on. One of the things that we always have to do is balance that tension between providing things that the entirety of the company needs without influencing or particularly impacting things that are very specific to a particular business.

And so as part of that, one of the things we set up was we just need a way to capture all these ideas. It's been really interesting because it started, you know, over time with like a little couple of things. And then as people's understanding of AI grew, as they understood where it could be applicable, as models improved, we started getting more and more to the point that now we're at that clip where it's approximately almost a thousand a year. And it's given us great insight into different things. The piece that I always like to think about a lot is, you know, a few years ago when we first kicked off this approach.

a little over half of the requests or ideas that people around the company were having were chatbots. And if I look at today, a similar percent is agents. And so it's also reflecting in the way, like, how are people thinking about improving the way they work? How do they are thinking about, you know, deploying AI? And also, it's actually a reflection back to what I was saying earlier in how things are improving and how those patterns are changing, because now we've gotten to a point, right, where you can expect AI to take action. You can embed it into a particular process. And it's good to see that change in mindset over time as well. Yeah, I use a Fletch reference for...

agents now. There's like a scene in Fletch where he plays an airplane mechanic. He's impersonating an airplane mechanic and he takes the name Gordon Liddy. Gordon Liddy. Who's I think somebody went down in the Watergate scandal. That's right. He walks up to these guys who were actual airplane mechanics, right? And he's got his uniform on and he's like, hey, what do you think? Like, you know, and explain some problem he doesn't understand. And he's like, ball bearings. It's all ball bearings these days. And I feel like that's agents. Like people are like, hey, so I want some AI. What should I do? Oh, agents. It's all agents these days. But they don't, they don't. There's two things that's interesting about that. Like, first of all, you go buy from Salesforce or somebody like a.

a warehouse full of ball bearings, right? Like agents and the mechanics are like, what? We don't actually need ball bearings. And the other thing that's most interesting about the analogy is that if you, even to an engineer, if you took two ball bearings and just on site alone. had them tell the difference. So call this like an agent demo, right? Like someone shows you a demo of an agent doing something or a vendor showing an agent demo. It's even for somebody sophisticated, you can't really tell the difference. But it has to be perfectly machines, perfectly balanced. The material has to be exactly right, polished, et cetera. And the difference is if you put a bad ball bearing that looks exactly the same into an airplane, that engine blows up. And it's very similar. Like these demos are not viable ways to discern the difference between point solutions that...

will work and point solutions that won't, which is so tough because if an engineer can't tell the difference by looking at two, then demos aren't going to cut it as far as trying to figure out what solutions to use. I don't know what your thoughts are on that. I think it's a very fair way to put it. And maybe I would break it down to almost like two principles that we've been trying to focus on. The first one is actually on training and fluency. I think it's a very fair point, right? Everybody's asking for agents so they fully understand what an agent is. What do you think is important, regardless of whether folks are calling it chatbots or agents, what I care about from my standpoint on an enterprise side is...

Do all of our employees understand what capabilities we feel are ready for use, in what way, in a way that's responsible, in a way that's safe, and that they have an understanding of how to embed that in their daily work or in the products that they build? I think that's the important piece. So almost the point is, they might call it ball bearings or not, but if they understand that, it has to be machined correctly. It has to have the right materials. It has to have the right precision when you manufacture it. I think it's important. The second piece then becomes into that precision piece. It's about trust. It's about responsibility. It's about doing so thoughtfully. And a lot of what we spend our time in from a process standpoint is embedding some of those.

some of that intuition into the way we develop either internal products or market facing products, right? Is have I made sure that a particular AI solution has a human in the loop and very critical points that could cause a failure that we don't want happening or that. you've gone through and checked for things like bias and tone of voice if you're using generative AI. I think breaking away from the name and saying, what are the core capabilities I need to make sure? And by the way, which ones do we think for very clear reasons are ready for enterprise user or not? I think has been a lot of what we focused on. And so to me, going back to the first piece, it's very flattering to see everybody kind of ask for agents. Underneath the scenes is if what people are understanding is a lot of the ideas have the right tools or the right pieces of workflows where we do have confidence, then we're doing something right from a training and enablement standpoint for our people. Yeah.

This episode is brought to you by OneReach.ai, the runtime for AI agents. Their GSX platform provides turnkey private architecture and tools for building enterprise-grade agentic solutions. Recently, I worked with the team at OneReach to create an agentic companion for the second edition of Age of Invisible Machines. So now when you scan any of the QR codes found in either edition of the book, an agent is there to provide additional insight using a custom knowledge base. There's also a newsfeed feature with updates on agentic AI as well as this podcast and a book resources skill.

that provides clickable galleries of footnotes from each chapter, all of it built using the revolutionary Generative Studio X platform. Learn more at OneReach.ai. We have a tough time getting our head around how fast things compound, right? And so when we think about some of this technology, we think, okay, agents that can help me do this task versus agents that build agents. And you're like, well, agents that build agents is compounding, right? Agents that do tasks is point solutions. And so you start to think, like, if you're going to attack the AI problem in an organization,

you can first principle it and say like, well, let's attack the things that can compound in this. How do we make AI compound in this organization? Which would be like, how about starting with AI that educates the organization on AI, right? Because now you're kind of hitting first principles. You're like, AI will make itself. It's the same thing as like robots that can make robots. If you have a robot that can make itself, then you have this compounding effect. And the other thing I think, if you're struggling with compounding, then you're really gonna struggle with the idea that these things can make a better version of themselves.

Because now you're not just talking about the sheer compounding effect of AI making another, an agent making another agent, but an agent making another agent that's better than the original, which really gets into like human evolution, et cetera, right? When we talked earlier, I really... grabbed onto that as you were talking, like this idea of first principle teaching, like you can't take a thousand meetings a week where people have questions on how they could use AI to do their job better. There's just no way that you can scale. If you scale you, right, versus scale them, then they can scale them. And I think this is a fascinating idea for anyone who heads AI to focus on AI for educating. That to me is a fascinating and evolutionary approach.

No, I think, look, I think it's a very fair point. I'll pick up on one thing, which is the first principles piece. Very early on when we established our enterprise AI and data team, one of the first things we did is we said, what's just the grounding principle from a strategic standpoint we want to take, right? How do we basically... give the rest of the organization some framing about when and where to think about using AI. And we landed on a framework that's quite simple. We like to think it's catchy, but it's actually really about focusing, right? So we very actively talk about using AI at MasterCard for making commerce more secure, for making it smarter, for making it more personal, and for making MasterCard stronger, right? And it's kind of very related to our core components. We think about our in-commerce. How do we drive commerce with our payments networks, with our value-added solutions? And what that's allowed us to do is to say,

This is going to keep evolving very fast. It's going to keep getting better. The way we use it will expand. But focus on deploying it in just these three ways or to make the company and the way we operate stronger. These are so many things we could do, but then go back to first principles. Just focus on these three things because it's a really clear way to think about when to apply it and where our more strategic priorities are today. Then when you say, how do you prioritize? Where do you invest? How do you prioritize? What you move faster versus slower? That very simple grounding becomes almost common language for us. Everybody's saying, well, this is about making commerce more secure. How so? Okay, let's take that forward. And if it doesn't fit in that framework,

could still be valuable, but at least you have a common way of approaching the problem of where do you deploy as it continues to get better. If instead of like looking bottom up, like how do we automate this stuff? You look at like, let's create a DNA for AI in MasterCard, right? Let's create a DNA strand that will be replicated across the whole organization. What are the core components that are shared in terms of AI? What are the core values of the company? How do we embed those in? And then with that DNA, then that kind of spawns each individual kind of thinking about it.

And some people would call that guardrails, but that's way too narrow to think about it, right? Because it's both guardrails, it's creativity, it's understanding, it's learning, right? It's not just guardrails. And this idea of creating a core DNA first and using AI to help you replicate that DNA internally is, I think, a really cool and unique concept. Yeah, I think that's right. And that helps you drive scale, almost to put it in a different way. The best people to think about where and how to deploy AI are the people working on different business problems.

We have many of those people working to support the ecosystem, to support our clients on that. And so part of the idea is how do we empower them? And then from a more enterprise standpoint, or from looking at the enterprise lens, how do I and the rest of our team focus on the very common problem, like the least common denominators that then help drive that scale? Yeah. Yeah. It's like everybody wants to do something with AI, but really in the beginning, it's like start at the top, which is know something. You need to know something before you do something. If you do something without knowing something, you're going to do it wrong, which is what we see constantly. As people say, agents fail. What does that mean? Agents start to do tasks before they are trained in the knowledge to safely do tasks. So because people go to doing, you think of like, we use Patagonia as a good example. Like if you hired somebody to work in a store,

you know and you're like okay great you're gonna sell wetsuits like what's the first thing you're gonna do are you gonna teach them how to you know, how to open the front door and how to greet a guest? Or are you going to teach them the knowledge of wetsuits and like what you need to know about wetsuits first? And then you put them on the floor to actually sell wetsuits. And we just know this intrinsically with people, but with AI, we tend to like think of it in the opposite. Like, no, no, I just want to get them out there on the floor selling wetsuits, but they don't know anything about it. Like they just sold a five millimeter wetsuit to this guy and we don't have a five millimeter wetsuit. Like we don't make one. Yeah, I think that's right. The piece of context of deploying it with purpose, of embedding it where it makes sense. And honestly,

having the human in the loop in the parts of the process that are still too sensitive, I think it's pretty critical. Yeah, yeah. Yeah, and I think having that knowledge in place is really important too as we start to move into this era where agents are going to have wallets. I don't think we're seeing a ton of it yet, but I think it's just around the corner. We're going to have agents buying things on behalf of people, people sending agents out to make purchases and really maybe even not paying that close of attention to those purchases. That really points to, I think, the key value prop for MasterCard, right, is trust. That's the most durable asset that you have. And so make sure that anything agentic happening in the organization

is strengthening trust and keeping trust going in a world that's like suddenly very shifty and confusing, right? It is. And we know there's interest in using agents to drive payments, right? And a growing amount of online traffic with respect to product discovery is happening more and more via LLMs. And so we know that the consumer demand is there. We know that one of the key pieces that was missing was closing the loop on that, right? You discover a product and then you just want to say, okay, I decided on X of the recommendations, go ahead and execute that payment of that transaction for me. We knew that was coming. It was coming rapidly. And so...

The first thing that we as a network and a provider of many commerce solutions have to do is we have to make sure it's safe and it's equally secure and it's equally seamless to have an agent execute a transaction on your behalf, much like it would be tapping your card or your phone or whatnot device on a terminal. It's really, really important. And so when you think about that, then you say, well, what are the core components of that is I have to verify who an agent is and that they have the delegated authority to do so, that the credential is right. I have to make sure that every party in a transaction is able to... understand that this is an agent and what safeguards need to be put in place as part of that and so at the very core we've been focusing much on that like how do we drive that interoperability we have things like you know an acceptance ring where it talks about what are the type of credentials that get exchanged as part of this if you're a merchant what is the low lift way to start accepting you know the type of tokens that are used by agents uh you know the core adjacent capabilities like having the ability to do agent sign up so that we only allow registered agents to transact on our network supporting many of the integration patterns as you've seen in the many protocols like that type of

behind the scenes work is actually quite critical because if we from our standpoint can enable that very well early on then that allows you to scale into moving to much more sophisticated transactions right today it's you search for a product you discover the one that you want agent executes that transaction but understandably so, many people jump ahead to, well, could it do multi-card checkout for me? Could it actually research a trip itinerary and book all the adjacent pieces and make decisions behind connecting the dots on this? We're all really excited about that, but that only gets there if the ecosystem moves through what I would call almost the base case, just being able to discover a product, execute that transaction, doing so safely, and doing so in a way that you feel and you see is equally secure, is equally seamless, is equally painless as using any of your other credentials today is. That is really, really important. Yeah.

The idea that at a logical algorithmic level, the transaction, you know, the moment the money transacts and all things are handled, like race conditions, et cetera, like the money is now irrevocably, irreversibly done. I know that's taking it too far. It's not quite irreversible, but generally that's like the moment that whatever was happening before culminated to that moment, right? Now we exchange money and there's a big responsibility on that moment because that's the moment that says, okay, are we sure? All the ducks are in a row.

Are we sure that that's the actual person that this transaction, these are the right parties? Is this the right amount? And have all of the elements that are their prerequisites to this transaction, have they all been met, right? Okay, like we'll make the transaction happen now. So it's kind of like in the chain of whether it's AI agents or not, this is like the most important moment, right? And I think as things get more complicated, because you mentioned it, right? I always like to think about buying

you know, airline seats. Because if I said, like, how much is an airline seat? You'd be like, what? To where? And then I'd be like, how much is an airline seat from San Francisco to New York? You're like, what time? What day? Right. Okay, this time, this day. Okay, how early did you book it? There is no such thing anymore as a price for an airline. Like, this is algorithmic and it constantly depends and it's complex. There's no static price now. And we're going to see this across the board on merchandise that's typically static price. We're going to see...

all of this variable pricing being built in based on demand, supply, who you are, what you can afford, how bad you want it. As we see this in travel, we're going to start to see this in buying groceries, right? I'm not sure. I'm not sure yet. I think it's going to get very interesting because while having buyer agents and selling agents does open the capability for a lot more price dynamics as a consumer.

I still have an expectation of what a number of things cost, right? And that usually drives some of my behavior. And so where we're more focused on is talking about complexity. Think about the different patterns that could emerge from just a transaction standpoint. There's going to be agents now doing transactions. Maybe they'll do a number of transactions in sequence much more rapidly than you as a consumer would, right? So the way the patterns look in our network are going to start changing. And again, part of what we focus on is how do I first help everybody in the ecosystem verify that the agent is doing a transaction that you gave it authority to do, that it's actually...

exchanging the right information with the merchant to ensure that you get a pair of shoes that you wanted or the plane ticket you wanted or whatnot. And then once that happens, live while the transaction is being routed, things like our decisions intelligence capabilities, how do we in real time understand a number of variables to say this is right. This doesn't fit a fraudulent pattern or it doesn't fit a set of relationships that point us to be concerned. And then the last piece, which again, things that don't get talked about a lot is as we manage these networks, there are rules. There are rules for when a transaction can be settled and what that looks like and what happens if something was incorrect or services or goods never tendered. And that helps protect everybody in the ecosystem. And extending those rules and making sure those frameworks are in place is actually quite important because back to your point, the world is going to get more complex and having that base understanding of

how do we all agree to operate with each other to make sure that at the end of the day, when you tell your agent to execute a transaction, that it's going to mean that you eventually get what you ask for and that the party that's doing selling is going to get paid for it. It's really important. I think it's an important point because I think on the surface, people might think this middleman handling the transaction becomes less important in AI when in fact it becomes like 10x more important. It connects with other things, too. We had a conversation with Joshua Ganz recently, who's an economist and a professor at University of Toronto. And he sees AI as just cheaper prediction, basically. And so on his model, at some point, Amazon gets so good at predicting things that they just start sending you stuff. I think he called it ship-then-shop, although it seems like the shop part kind of falls away. But Amazon just knows you well enough that it just starts sending you stuff. And so in that scenario, if MasterCard's in that ecosystem, too, then the trust is this shared, passable thing that almost has to be agreed on by different entities within an ecosystem.

Yeah, I think, look, trust for us is the currency of innovation. If you think about MasterCard, you think about, we safeguard a lot of trust. That is core to what we do in everything we do. And as the world gets more complex, as some of these patterns get more complex, we find it to be part of our role to make sure that the ecosystem can operate in a trusted way. If you think about that example and you take it logically, again, I try to use myself as a consumer, as a grounding mechanism very often.

I don't know if I'm always going to be comfortable with something like ship and shop, right? And so for some things like paper towels and toilet paper, if I could have an agent that learns how frequently I need it or I have some more automated way to do it without having to tell it, send it to me every couple of weeks because I behave differently every day, that would be great. I would be very comfortable with that, but I would only be comfortable if I can trust it to have a number of controls so I can delegate some authority to the agent. I can say, I'm willing for you. to more autonomously send me paper towels and toilet paper, but don't do it more than X frequently or don't spend more than this much money and only do it via these merchants. And actually those controls is very important. And I want to make sure that my agent, not just has card rails, but that can only transact with parties that I think that are going to be verified, that are trustworthy, that kind of operate again within that same kind of rules. I as a consumer want that because even for somebody like me that's spending every day at work thinking about how do we push the way we work forward? How do we innovate in what?

you know, millions and billions of people use every day with AI, I still think about like, what am I at a core truth very comfortable with? And where does my trust lie versus what am I just not ready for? Or what is true for me to be ready for? And if you think about it with trust, and this kind of gets back to our earlier point, trust is about the knowledge that the system has, like everything you just said. Like, let's say it has the knowledge of your smart thermostat, right? Let's say that's part of its knowledge. Now you might think, well, what does that have to do with toilet paper?

But if the thermostat's been at 50 degrees for a week, you're not home. And if you're not home, you're not using toilet paper. So now that the system knows you're not home, now before it goes and buys the toilet paper, it's going to confirm with you, like, hey, I don't think you're home. You want me to skip toilet paper work this week? Now, what just happened? Was it guardrails? No. Was it trust? No. It was knowledge. It needed to know and have full context. Without that, you're going to have it ordering toilet paper every week, and then you're going to have a closet full of toilet paper, which might be fine if there's another epidemic. It could also be a fire hazard, though, you know?

Yeah, fire hazard. And I think this is important because these, what I would call dumb algorithms, right? These like basic algorithms that so many companies have tried, like automated. I remember they had like, Amazon had, you know, these buttons that they would put and then they would auto order. And then like my kids, when they became toddlers, found the button. And I had like, yeah. And I had put it under the sink. I thought I was really smart. You know, I put it, you know, where.

where you would notice that you're out of dish soap so that you just press the button when you see the dish soap getting low. And yeah, I had like 100 cases of dish soap show up. So again, not to say my kids don't have knowledge, but they did that too. And I'm likening agents today as two-year-olds, right? Like, yes, agents make sense. A two-year-old agent that's in charge of B2B helping you shop a B2B software solution for your company doesn't make sense. What makes it a two-year-old agent? The knowledge it has. How basic is the algorithm it's operating on? And so Josh, I think that ties into this whole knowledge management is also something that's...

that people address on a point solution basis. Like we think of it as a document repository versus like, how do you actually manage knowledge in an organization in an AI first sort of way, in a way that's not about how we used to do it. I think it's pretty foundational. And I remember when we originally were chatting, I shared one of the most common patterns of development we had early on when we really, when Generative AI really started permitting across the enterprise was that, was start with getting a knowledge management solution in the hands of a team or employees. Because that then helps you start figuring out where is the AI tools you have at your disposal? Where do they behave like a two-year-old? And where are they very impactful? So let's take the example you said.

make a decision on a B2B software solution? No. But you know what happens today for a number of material decisions on that? You do an RFP or an RFI. And that requires a lot of information. And typically each party on the issuing side of an RFP is looking for different things. I will be looking for different answers than my security counterpart, than my technology counterpart. And so actually I would use an agent to help me answer those questions. I would not today use it to make a decision, but I would definitely streamline the way that I do the evaluation because I have something incredibly powerful at synthesizing that knowledge.

but where I actually want critical thinking, like critical human thinking today for the decision-making process. Yeah. And 150% on that, 150%. Because what we're talking about is, is again, back to learning. Like first principles, if we teach how to select a solution first, then you don't need an agent to select the solution. We need an agent to help teach people how to select the solution. That's right. That's where you begin. And if you teach an agent to select a solution without...

teaching people or the agent itself the knowledge, then you're going to get the wrong solution. It's so obvious when you think about it. We never do this with people. Or think about the way you develop products today. Think about your average product manager. They spend a lot of time at the very early stage of the funnel on knowledge management, right? Talking to users, observing behavior, and synthesizing into requirements. Well, what I really want to get out of a very talented product manager is not the procedural or mechanical step of consolidating feedback into...

stories that then become requirements i want them to think about what do you do about those requirements what do you do about how do you actually solve the need which is continues to be an extremely hard problem and that's where you really see very talented product managers shine versus the normal ones right yeah so it's all about shifting the way you consume knowledge or frankly how you deploy the tools because in when you are now accelerating the way you make requirements into things like okay making a decision can i augment that product manager by coming up with more ideas or more rapidly prototyping so they can test their ideas right I get very excited about this topic, but it's so powerful when you get bored and fall asleep, when you can take that step from just knowledge management, say, oh, because I have much of my knowledge, I will now think about step X differently so that I can, I, the human can spend more time critical thinking or evaluating something, which is, you know, that influencing others, communicating a story very clearly or things that continue to require, like, I think are requiring more and more of a premium on people than they use.

100%. And I think this idea of things being boring, you know, if things are boring, they're good in AI. If things are exciting, you're probably doing things wrong. And if knowledge management is boring, then that's what you should lean into. We're in this paradigm of knowledge with these systems, like what I would call the chatbot era, which is you ask it the questions and then it gives you the answers. So you need to know what questions to ask. We took our best-selling book, Suck It Into an LLM. This is like years ago, right? And we're like, aren't we smart? Now you can ask it questions. And then people would sit there with like, why don't I have to ask it?

Good point, because that's not how we consume knowledge. I always liken it to the classroom. Like, oh, okay, you're now a teacher at MIT. It's your first day. You arrive in the beginning of the class. There's a bunch of students just eager to learn. And you go, okay, any questions? Anyone? Like, I'm a chat bot. Just ask me what you want to know. And they're like, what is this, right? These things need to be proactive, not reactive. And this AI-first approach to knowledge management is about not just like it having the knowledge, but it proactively sharing the knowledge within your organization.

and curating it and driving knowledge down instead of waiting for knowledge to be requested. And Josh, I think we have some kind of demos. I can kind of show you what it looks like for us right now. What you're looking at right now is actually a knowledge model of pretty much every episode of the last season of this podcast. And it's broken down into tags, which are topics that were discussed, areas of interest. And then you have these notes, which correlate to specific conversations or moments of conversation that occurred in different episodes. Yeah, these are like canonical ideas. Very cool. And so when you look at them together...

um like when i first loaded the information into this model it kind of had more of an orb shape and like a lot of the ideas were sort of disconnected but then what you actually do is you kind of go through and verify each one and as you're doing that you're also deciding like what other tags it connects to so like is the knowledge in here correct and what other pieces of information in this knowledge model are related to that knowledge and so what yeah like what you're doing is kind of going through and you're saying that this knowledge is essentially ready for consumption. And Rob, you wanted to talk a little bit about the composition of this too, right? Yeah, the idea is, imagine Patagonia. Imagine you have a knowledge model for every product, right? So wetsuit, a specific wetsuit would be a knowledge model. And then within that, you have these categories like weather, materials, sustainability. And then under those, you have ideas that are canonically, like the thickness of the wetsuit, where you're not going to have that replicated, so you don't have an overfitting problem. It's like, yeah, that's canonically correct. So this is why it's not your document repository that's a mess of...

from a canonical standpoint. There's no truth. This is now a source of truth of knowledge, right? With a history in each one of how that knowledge changed over time. Now, what is really cool is you look at it like a map, and I don't know if you're familiar with the traveling salesman problem, but it's like this idea that says, okay, well, this is a traveling salesman problem. Now, we're gonna teach you everything. To become an expert, you need to learn everything that's in this map right now if you wanna be a full-on expert. So what are we gonna do? The first thing we're gonna do is we need what you know. So we go and we make a digital twin, a learning twin.

of all this knowledge, right? And we go, okay, this is you. Now this now represents what you know. Then we do an assessment. We go through and assess what you know and what you don't know. And we light up all the things you know. And now we have a gap analysis of what we need to teach you. And then we do traveling salesman. We go, what do you need to learn next? We don't try to map out a curriculum for everyone. This is your curriculum for what you need to learn next. We consume that. You learn it. We get feedback. And then the next thing you need to learn. And so you're traveling through this, learning what you need to learn. And what it accommodates is A, where you start. It's not the curriculum that's for everyone starting from the beginning.

It's a custom curriculum for you based on what you know. And then the second piece that it's doing is as this stuff is dynamic and the information is changing, you're picking it up as you go because it's not creating a curriculum that two years later you're teaching the same curriculum even though everything's changed. So you have accounted for the dynamicism of the knowledge that's now picking it up as you go. So now you have this GPS for learning which says here's where you are, point A. Here's where you want to go, point B. And turn by turn, we're going to get you there. But at each turn, you could go a different way and we'll recalculate the route.

And so this is all just happening, and it's canonically correct, so we're not going to teach you the same thing three times, which is ultimately the objective. And this is AI-first learning versus the opposite, which is somebody goes, asks a bunch of questions about what you know, puts it into an LLM and says, based on this, create a curriculum, and it just generates scrap, right? The piece that also jumps out at me from this, and it's something I spend a lot of time wrapping my head around is... you're not just talking about the capability you're talking actually about a almost like a cultural way of working shift and i think that's one of the things that is quite underappreciated about this whole transformation like it requires teaching people to change how they work to think differently and that that is at an enterprise level for us like that's a cultural that's like a big change right it's cultural shift it's it requires people to think differently to think about where in your day-to-day are you going to adopt a tool like this that all of a sudden right will help you learn better will help you improve in a number of things that you do but requires you to engage differently you might not think about training the same way or you might not approach the same way and that's me on top of the fact that it's really creative about tweaking the knowledge base and all this is

you're also changing the way somebody's going to approach the whole concept of training or readiness as a whole. This is just like one of many front ends you can now put on that knowledge. Like the whole idea is now that I have this knowledge on the wetsuit, I can deploy agents that can do things like handle supply chain, right? So now you take that knowledge in the center and then you orbit the agents around that knowledge. But without that, you've got these agents with no core, right? With no brain, aimlessly trying to accomplish stuff out of context. And so this is now like, okay, now we'll take that and create an agent to teach.

Very cool. This agent is ready to teach us about everything basically that was discussed last season. And like Rob described, first you can go through and read through the different topics that were discussed and kind of mark yourself. Like, am I an expert, a novice? How much do I know about all this stuff? So then the system kind of knows where to begin.

That's where you are. That's point A, yeah. And then here, like, what's your objective? Like, am I curious? I think for a lot of people in enterprise, it's going to be upskilling or mastery. And then, you know, you can put in a specific goal. Then it creates a curriculum that's tailored to what you're trying to do and what you already know. And the idea there is that, like, you know, we were kind of thinking in terms of our book, right? It's like this thing that you read from the beginning to the end for most people.

But, you know, there's a world where like what I really want to read about are design patterns because I do a lot of design. So I want to start there. And so if you kind of like lead people to the information they're most interested in, then you can start to build out that. Yeah, the knowledge. And I think that's really important. Like once you start to understand how much everything is learning, you realize like, I want to buy a wetsuit. Which one should I get? That's learning. What do you know about wetsuits? What don't you know? So now the sales process is learning is a class, right? If all of your folks are looking at AI products, each one of those, like each vendor should be providing.

this, right? Like, what do you know? What do you want to do? I'm going to teach you the parts of this product and how you would do it with this product. Now we're talking about educating people on how to make choices on which AI products to buy versus asking you to help them pick one. Yeah. Yeah. And you over time, right? As your catalog changes everything, you'll just continue pushing on that front. You'll continue teaching people about the next thing and it makes a lot of sense. Yeah. You can imagine an organization where you could have a knowledge model like this that's designed specifically just to get people up to speed on agentec AI and figure out like how they'll make use of it. But then throughout an organization, you might have all sorts of

of organizational aspects that you really wish everyone knew, right? Here's how we do this specific thing. Everyone needs to know this, but everyone can kind of learn it in the way that will work best for them. Yeah, and I think the most important thing in that, what you just saw, is it's not building the 10 steps or the 100 steps to be an expert. It's just looking at the next one and then recursively taking what you learned so that if you go learn something somewhere else, right, it can accommodate that.

You don't have to also learn it here. It can say, okay, you already learned that? Okay, we'll skip that and go over here. And then the other thing I really like about this and how it changes the paradigm of learning is we're so used to teachers being the delivery mechanism for the teaching, right? They'll communicate the learning. They'll be the vehicle of the ideas. But if we think about teachers as a facilitator...

not necessarily the deliverer of the ideas, and then we look at this as being a facilitator of the idea, then it can see an idea that you need to learn next, but it can go find a good YouTube video on it to teach it. So it doesn't have to facilitate the teaching itself. AI first isn't about technology problem. It's about rethinking. You said it earlier. Yeah, you said it earlier. It was like rethinking how you do business, how you think. Yeah, that's the most exciting part of all of this transformation is how do you rethink something to get...

to more value or to really, you know, focus somebody's time on the pieces that are going to be most impactful. Or for example, back to training, where do you focus hands-on learning, what you have to do versus, you know, asynchronously capture ideas, et cetera. I think AI has a lot of promise about rethinking some of the ways we've done things. But again, maybe back to the beginning of our conversation, what is the point of teaching is, well, I want somebody to learn these concepts so they can be effective at doing this or they can be great citizens or whatnot. Therefore, what do I need to make sure the person is getting out?

I think it kind of goes straight back to the beginning, which is like, what's that end state that you're looking to get? And then how do you deploy slightly different tools, some very advanced now to get to that outcome? Yeah. Yeah. Yeah. And I think too, just this idea that that with trust being such a such an important feature in anything really AI related, especially with commerce, you realize that knowledge is is almost inextricably tied to that, right? Like agents can't operate without knowledge. Humans have a hard time operating effectively without knowledge. And one thing we didn't show was that that system, that expert had the ability to go out and learn every day.

So as knowledge is changing, it's proactively meta-learning, right? And that could be talking to individuals in a company. That could be searching the web. That could be- Yeah, additional files. Listening to the podcast, right? Listening to the transcript. Yeah, I actually saw a good site. I don't remember the name of it, but they aggregate all the transcripts of the top 5,000 podcasts. I'm like, okay, I'm going to make an expert out of that. I'm going to consume those by different categories like AI. And I'm just going to have it every time a podcast comes out, I'm going to have it consume.

that knowledge. And what's cool about this idea of compartmentalizing the models like this is if the subject matter or the transcript has nothing to do with wetsuits, for example, nothing gets added to the model. So we enter this world where like, don't give me the summary of a meeting. Tell me the pieces I need to know. Give me the transcript. I'm going to feed it to my experts. Each one's going to pull out different pieces that are relevant to me. I don't need your...

two-year-old agent, you know, summarizing this, I need the ones that learn. So if we talk about wetsuits, that will get added to my wetsuit expert. The other expert might be healthcare related. It's going to pull those. And when we were talking about my vacation, that's not data that's going to get consumed at all.

And you're going to want human experts, like to the human in the loop piece, I think you're going to want human experts up and down this stuff, making sure that what's being taught is accurate because transcripts of all the most popular podcasts doesn't necessarily mean true things. That's incorrect. Yeah. That's a good point. Fact checking. Which is like another thing that people call it. They'll call it fact checking or guardrails. But in the end, it's just knowledge management. It all rolls up to, is this true? You know, is this the right person? Did they mean to buy this? Is this the actual price they agreed on? It really all rolls up to knowledge.

That's right. It's knowledge management. It's context engineering. It's framing the right pieces and then developing algorithms, processes, approaches that kind of drive that little trust in a non-deterministic set of tools. drives a consistent alpha. I'll have my MasterCard agent that knows all of my context for what I buy, and if some weird transaction happens, it'll be like, text me, wait a second. I'm trying to order a toilet paper again. I don't know if you're ready. 15th packet of paper towels on his web. Yes. Well, cool. This was great. Yeah, super fascinating. No, please. It's lovely to have the conversation and I thought it was super engaging. Love what you guys are doing. It's awesome. Thank you.

Thanks for hanging out with us on Invisible Machines. Remember to like, subscribe, and or follow us wherever you get your podcast media. Thanks to OneReach.ai for their support and much gratitude to the many people working behind the scenes to make this podcast great. Until next time, Awagi.

Frequently asked questions

How long is this episode of Invisible Machines podcast by UX Magazine?

This episode is 51 minutes long.

When was this Invisible Machines podcast by UX Magazine episode published?

This episode was published on May 21, 2026.

Is there a transcript available for this episode?

Yes, a full transcript is available for this episode. You can read the complete transcript on the episode page.

Can I download this Invisible Machines podcast by UX Magazine episode?

Yes. Use the download control on the episode player to save the publisher-provided media file.