How JPMorgan Engineered a 30K AI Agent Economy

Over the course of those first few months there, we went from zero users to 250,000 users. There was actually a little bit of a competitive angle to it when your neighbor has it and you don't. And I think that gamification was actually helpful to turbocharge the adoption. Today, one and two JPMorgan employees around the world use it nearly every day. - This is Ventubete Beyond the Pilot. Enterprise AI in action. I'm Sam Widerving. I'm Matt Marshall. Today's episode is presented by OutShift by Cisco. Cisco is emerging tech incubation engine and driver of agentic AI, quantum, next-gen, infra, and beyond. Today we're talking with Derek Waldron. Derek is the chief analytics officer at JPMorgan Chase. While many enterprises are just starting their AI journey, JPMorgan began building their internal platform, LLM Suite, over two and a half years ago. That's long before most people had even heard of chat GPT. The result is one of the largest enterprise AI rollouts in the world with one and every two employees using their tools daily. For us, what makes the story so compelling is their early contrarian insight that the models themselves would become a commodity and that the real challenge and the defensible mode is in the connectivity around the system. Welcome to Beyond the Pilot, Derek. - Thank you very much. Nice to be here. You started this journey two and a half years ago. What were some of the core insights that made you focus on building this unified internal platform instead of just sort of waiting for vendor solutions to come along? - Yeah, so it was certainly early days and I think like everyone, we were also trying to figure out exactly how this technology was going to be used, where it was going to have impact, what the strategy should be around it. But I think very early on, we had three insights, which became principles of our strategy. One of the principles, which I think many people have realized now, is that we have to actually distribute the technology into the hands of the employee base at large. It's a general purpose technology. It's very useful in a wide variety of situations, but unless people get access to it, you're squandering the opportunity. The second insight we had was that no two job functions in a large enterprise are exactly the same. Even within a team like operations or like with sales, the day-to-day things you do are different than what your neighbor is doing. And that can't be solutioned from the center. If the center, an AI team, is trying to build out individual solutions for the organization, they just can't scale quickly enough. And therefore, the only way to to crack that is to be able to give very powerful, reusable building blocks and capabilities, that then the individuals around the firm can be used to build solutions themselves. And the third principle was the realization that the actual long-term bottleneck for driving maximum value from this technology was not going to be about the model. It was going to be about how the technology connects into the technology estate and data and process estate of an enterprise. While there was an arms race and new models were coming out, and there was a lot of speculation as to what model was going to be the best and who was going to win, we dissociated from ourselves from that and said, it's sort of irrelevant. Somebody's going to win or multiple people are going to win. There's going to be great models out there. So let's have the thought experiment. At some point, a very powerful model called superintelligence is going to show up on the doorstep. And then the problem statement is, how do you use it? How do you put it to use? And the realization was that even if superintelligence were to show up tomorrow, there's no value that can be optimally extracted from it, if that a superintelligence can connect into the systems, the data, the tools, the knowledge, the processes that exist within the enterprise. And that's actually a long pull in the tent that needs to be solved for. So all of those three principles was what gave rise, then they said, OK, we're going to build a platform to solve all three of these things. We want to get the technology and the hands of the population. We want to tap into the incredible innovative culture that J.P. Morgan has and tap into the innovation that the employee base can bring. And then third, we want to solve in a very, very strategic way connectivity across the enterprise to supercharge all of this. I'm really curious. What was the hardest part of that sale back two and a half years ago? And especially the concept of sort of like, oh, we should give these models our data, right? Like a lot of people will, like, ain't no way we're going to let any of our data touch any other sort of AI model. How did you go about selling that? Where was that difficult in the organization? Was that easier with client advisors and harder with other parts of the organization, et cetera? Yeah. Well, first of all, we didn't give our data to large language models. Data privacy was the first and foremost consideration in all of this. So part of also the reason that we wanted to build a platform ourselves was that we understood then the lineage of all of the data because security was the primary concern at the time. And then in terms of how we sold it to the organization, we didn't actually at the time share the long-term vision. It would have been maybe too much to digest and too much to sell given the nascent of the technology. This was the strategy we had in mind. But what we did sell, first of all, was just the first of those principles getting the tools in the hands of the population. We didn't force it on anybody. We built a great platform. It had great capabilities. And then we invited people. What we saw over the first few months of launching it was it turned into a viral uptake. So people wanted to get access to it. It was cool to have access to it. And I think that gamification was also, just like we see in so many other technologies, it's actually helpful to turbocharge the adoption. In addition, we had a lot of marketing education around it. So you couldn't help but walk in some of our headquarters offices and see advertising campaigns around what the AI can do for you and how to use it in new features. We would regularly be talking about it in town halls. Some of our CEOs would be promoting the use of it. So it was really a sort of a full court change management effort, but over the course of those first few months there, we went from zero users to 250,000 users. And as you pointed out now today, one and two JPMorgan employees around the world use it nearly every day. Derek, I want to go into your own background. You have PhD in computational physics. You spend time at McKinsey, I think, for what, 10, 12 years focused on strategy, transformational strategy. And then you joined JPMorgan and it was quickly focused on strategy in the CEO's office, my understanding. Can you walk through how your experience was brought to bear on this key fundamental insight that you shared, which is assume superintelligence shows up on your doorstep and then everything float from there? Can you take this back to that experience that you brought? Yeah, I do think that at this moment in time, given that AI is deeply technical in nature, but in order to gain the value of it, there's also this very significant portion around organizational change management and transformation of processes. I do think that this actually now draws from the breadth of experience that I had. Obviously with a PhD in physics, I have the technical depth to be able to understand how the technology works and how solutions are made. But my time at McKinsey and my time in strategy was also where the bulk of my career has been spent was I think very helpful because it also taught me how you lead change and transformation in an organization. And that includes all aspect of thinking about reengineering, platform builds, the change management aspects of it, the organizational change aspects of it. And so I think that's just made it particularly helpful then when we started this generative AI journey at JPMorgan, there was some certain pattern recognition that allowed me to probably make some helpful, helpful choices of how to go about this. - So let's talk about the platform, LLM Suite. It got some great publicity, right? It vaulted you to I think the top of the AI evident index which tracks innovation in AI across all banks. So you're number one, you've got a big write-up in the Washfree Journal in February of this year talking about that success. You started by identifying common patterns like RAG and summarization as part of this suite. How did it evolve from a suite of tools into what you now call this connected ecosystem? - Yeah, the way that it started was basically solving for the distribution of just large language model technology to basically get it anchored on everybody's desktop within a hosted platform. Then that became the landing ground. But what we were also beginning to identify is in early 2023, we did countless ideation workshops with businesses, with operations, with technology, with probably thousands of people around the bank where we were demoing the technology and brainstorming, how could this be used? What type of novel solutions could this be used? And what we very quickly realized was that first of all, there was a ton of different ideas. You couldn't go into another meeting without another dozen ideas coming out. But in addition, they ended up being very largely restatements of the same type of solutions again and again and again. So one of them was RAG, as you said. People had knowledge and knowledge can be represented as policies, it can also be represented as a document that may sit in a deal room team if you're invest in banking, or maybe a repository of past information from sales. Everyone has their own knowledge expressed in digital form. And people wanna be able to use AI to query that and analyze that, so that was one pattern. Another pattern that emerged was being able to bring conversational questioning to structured databases. So text test, disequal, and there's commercial offerings for that, but basically automating the process of traditional database querying. And then there's other patterns that exist where especially in a bank, so many people deal with documents all day. Some of those documents are very marketing style, some of them are technical reports, research reports, some of them are our pitch books. But having a very, very powerful capability to be able to do document ingestion and analytics comparison, these types of things also was recognized as patterns. So we identified maybe half a dozen of the most prominent patterns. And we started to solve those. And then that began to allow us to move beyond just an LLM on people's desktop. To now we've actually got packaged, very flexible reusable components, which people then can configure in their own way to bring solutions to life. And that's what really turbocharged then a lot of innovation. Because people started to be creating AI assistants to do all sorts of things around the bank. Far more at scale than what we ever could have done centrally by trying to build out these solutions from a central team. - The connectivity, right? Rags is only as good as your data that you put into it. How do you sort of decide what data that goes into it? And my guess is that that was a progression. We're talking about something that's taken you two and a half years. How has that evolved over time? - Yeah, so the journey was like this. Once we started to solve some of the fundamental patterns and a lot of the patterns were about helping people to do question and answers, summary activities, etc. Then once that problem was solved, then people began, we began to encourage people to think about it, but also people began to realize, well, I don't just want to solve this part of a process. I won't be able to solve the whole process. And if you take a look at the types of things that people would do in a bank, an investment banker, for example, would regularly go to for a particular client, go to news, check the earnings release, do web research, pull from research about a particular client and then do a briefing note. A risk professional may do something for the same client, but it would look a little different. There they might look at news and earnings and look at scraping for troubled warning signs, but then they'll go to their risk data analytics database and start doing exposures and analytics, etc. And so what happens is that when you realize what the processes that people do within a bank, you've got to take a longitudinal lens. And more than often what that means is interacting with multiple different systems. That's especially the case in a bank, any bank, where you have a lot of different large tech data state because of how the banks grew up. And therefore you have to connect the AI into all of these systems. And so the connection problem is not just about bringing the data to your LLM or connecting it with a knowledge base. We realized we need to take this in a very holistic way. So we talk about the AI-connected enterprise as being one where we put AI in the center. And over time that AI will just become more and more and more powerful. It needs to be hosted at the center of a platform. But then that platform needs to have ubiquitous connections where people can interact with it directly with very sophisticated documents. It can be connected into knowledge stores that are provided. It can be connected into firmware structured data systems. It can be connected into applications like CRM systems or HR systems or other types of systems like that. And then it needs to be equipped with its own tools to help analysis and preparation material. And once you've got that AI-connected ecosystem at scale, now you have the basic ingredients that almost any type of job function, if they really think about what they do on a day-to-day basis in a longitudinal process lens, they can leverage the AI platform to be able to run these types of processes. And that's the vision where we're going. So we had that realization a couple of years ago and we built the platform around the concept of this type of ubiquitous connectivity. And then in particular, last year has been very much about then hydrating that with connections across the system. So we had connections into our trading systems, our finance systems, our risk systems today. And we continue to just add more and more connections now by the month. This series is presented by OutShift, Cisco's Emerging Tech Incubation Engine and Driver of the Internet of Agents, an open, interoperable internet for agent to agent collaboration. Learn more about the Internet of Agents and explore how agentic systems are the future at outshift.scisco.com. From a technical point of view, I can understand that you've got these real-time search things to get specific news or something like that. But what about the data that you had to reformat perhaps for vector stores or knowledge graphs? My guess is that J.P. Morgan probably didn't have a lot of use for vector stores before RAG. Is that something that you consciously decide, okay, all these knowledge bases, we're going to turn them into vector stores. And these ones we're going to go for a knowledge graph once you sort of test that sort of out. I'm very curious to hear that side a bit. Yeah, so we didn't make those decisions centrally. We built the capabilities. So the workhorse was a RAG system. We're actually in our fourth generation now of a RAG system. So it's continued to uplift in the capabilities to do more and more. But we made that capability available. And then we invited the whole population of J.P. Morgan to identify the knowledge assets that they thought were most helpful. And then they would then connect those and all of the mechanics behind the scenes were made invisible to them. So you've got four generations. What change? Can you give us sort of insight into what change? Yeah, exactly. So the very first release was maybe the most basic of just your RAG pipeline in a vector store. That was largely an AI pipeline which needed to be deployed in applications. The second generation was then where we democratized it. We federated available to the entire firm so that now people could contribute their own knowledge stores and set various access provisions around that. We call that knowledge-based connect. The third generation was when we began to realize that not all knowledge is actually created equal. So a business may have a set of documents that house a lot of the knowledge that it has. But actually, the types of questions that people asked, we realized that there's hierarchies of information. There's some information where you actually don't want to go to an underlying source. You want to script it answer, a frequently asked question. You want to verbatim. That's very important if the language matters. You don't want a degree of freedom. So that needs to be now at the top of the hierarchy. There's a second hierarchy of information when it needs to be evergreen. So you don't want to take that from just a vector score because then you're putting the burden on a knowledge store manager to make sure it's perennial. So you want to be able to carve out a pipeline that goes to the authoritative real-time sources. Then next in your hierarchy are sort of knowledge and information that's contained in documents that stands the test of time. And then your last of your hierarchy is information that may be more temporal. So it becomes less relevant as time goes on. And you have to bring that lens to it. And if you actually think about it, that's a hierarchy meaning that when a question is asked, what you want to do is start from top to top to top to bottom. So that required quite an uplift in the pipeline. And we figured that out by working with teams around the bank and understanding what they were trying to do and what they were unable to do and where they were having challenges. So that was the third generation. The fourth generation was much more around multimodality. Most rag pipelines started out as just text only. But if you think about so many of the information that exists within financial services, reports with graphs, or company pitch books that are very marketing-like, you need a visual pipeline to that. So that's now where we are with that. But we have a pipeline. It will continue to get more and more and more powerful. A very, very sophisticated document ingestion on analytics pipeline is table stakes to bring AI to life in the context of an enterprise, especially a bank. Derek, you talked about focusing on two core models, open AI and then Thropic, right, as a way of simplifying. And it's also reflective of your philosophy. It's not the model, right? A lot of companies we've talked with have a leaderboard of 50 models and they've got a sophisticated router and they're using different models for different use cases. Can you walk us through why that was so critical to focusing on this bigger problem of connections? And then maybe just a follow-up on the ramifications of that, you have something called Doc LLM, right, which is this layout-aware, generative model design for multimodal document understanding, specialize or customize to your workflows. How do you avoid, with your homegrown solutions, being behind the curve, right? And not being able to say use Google's pretty sophisticated Google-aware, I mean, document-aware technologies. Yeah, well, first of all, I think over time, we'll bring more models into the estate. It's diversification is important. Some models are good for some things, some are for others. So choice is nice because then people can pick the best that they, for the task at hand. But generally speaking, if you just have a sort of a smaller set of the most powerful models, then that's good enough. And then more of the solution, as we've already described, is really about the connections and how it works together in the ecosystem. We've always had a hybrid approach from the day one of both by and build. So the platform that we built was an aggregation concept. Where what we wanted to do was we said, we want to be able to build it with innovations that we haven't hydrated. But if there's good capabilities that exist out there in the industry, we'll bring them on board. And today it's a blend. Doc LLM was an innovation that came out of our research team. It was a really nice innovation because it applied transformer technology to visual forms. Think like a check or something like that. Where now you have spatial awareness. That was a problem that was very particular to financial services. And we didn't expect that a third party would come and solve that anytime soon at the level of rigor that we needed. So that's why we made that. And then keep your legs like that getting to great it in. But building capabilities and maintaining it also, of course, costs ongoing time. And so we're always taking a critical lens, recognizing at some point, we will just shift to third party solutions at the time that makes sense. You have the surprising viral moment that hasn't really been as well reported when you launch the personal assistance and kind of allowed individuals with the org to build these agents, which you're calling personal assistants, I think, 30,000 of these things have kind of bloomed. What was it about empowering individuals to build their own tools that tapped into this entrepreneurial spirit and created this blossoming? Yeah, so it's correct. A few months ago, we launched the capability of personal assistance to really empower the individual users to be able to customize an assistant with a persona, instructions, and roles, and available tools, and available knowledge. It was largely what we thought was going to be a convenience feature, because so much innovation was already happening around the bank, where people were building good prompts, effectively, what they were doing. But we wanted to give them a little more power and convenience rather than have to maintain prompts and a prompt library. We wanted to say, let's give them a nice capability just to encapsulate it as a more friendly thing, which we'll call an assistant, and we'll give them a little more flexibility that's not just about the prompt, but some other aspects of how it's used. And we've built that out over time as well. There's more and more capabilities of things that assistance can do. But largely, it started out as just a convenience matter. What we were surprised by was just how viral it was, because the organization now interpreted that all of a sudden, they weren't just designing prompts. They were building assets. They were building assistance. People were sharing the assistance they built almost within just a few days. Many of our lines of business started to create assistant web pages, where people could post and share some of the biggest innovations. And so that was really then, I think, eye-opening for us that there is this deep, deep rooted, innovative population with J.P. Morgan that if we can continue to equip with really easy to use really powerful capabilities, they can probably turbocharge the next evolution of this journey. If someone's building that, is that more than just a prompt? Like, can they select what sources of data can be pulled in by a rag? Do they have tools? Can this be hooked up to an email sort of system to automatically send emails or things like that? Can you give us, perhaps, some sort of tangible examples of those? And then also, it's fascinating that you've built this almost like we would refer to it as an agent economy in there, an assistant economy in there. It's sort of like a trading system almost of like, OK, we will develop this. And I'm curious for, as you leading this AI team, what do you learn from all of those things as well? So it is more than just a prompt. You can attach to it connections to systems which are already within the platform, connection to knowledge within the platforms. And then there are tools that are available, like data analysis and pitch book creation and these types of things. So whatever is available in the system is available, provided you're entitled to it. And that's a very important point, which is that-- because we built all of this from the ground up, we were also able to build in a way that was very, very integrated and compatible with our core entitlement systems and other good governance mechanisms. And that's a very important part of it. So when an employee wants to build something, they are only available to use what they're already entitled for. But within that, they have flexibility. The second thing that I would notice that we call these personal assistants, because these are for individual use only. But we do have the means by which then a great personal assistant can be created into more of a team or a firm wide solution, but that does then go through other types of governance mechanisms. And we've always had a very robust framework by which AI solutions go from ideation through into production along with all of the requisite risk reviews that necessary. So we have the means for that. And so what personal assistance I think is done is it's really expanded the front end of this innovation pipeline. Where now these assistants are being built. There's a lot of excitement around it. People are comparing. The best of them are also being identified. And those that are going to be promoted then into more less skilled solutions that are available than the firm. And in all of this, I think is your last question was around what role are we playing? How are we interacting with this? I think that what we've observed in particular in 2025 is that we've landed on an innovation flywheel, if you will. Where what happens is now is individuals around the bank have access to very powerful capabilities in building blocks. And that gives them ideas of things that they want to do. From there, they come and try and build them in some of them flourishings and scale. Some of them highlight that there's actually gaps inside maybe the connections that are available. We have a team that surveils all of this and surfaces up what those gaps are. And then we have a process to triage those so that we can solve them centrally. When we solve them centrally, we not only solve those particular individuals problems, but we also then expand the capabilities, which then creates more ideas, and we're in this flywheel of effect. I think the earliest days of that, even as much as a year and a half ago to two years ago, I had a realization for myself, which was that when I found myself calling someone for some information to ask them a question or emailing someone, I would immediately pause and stop and say, hang on, what's missing from my own personal AI assistant ecosystem? That means that I can't ask an AI assistant to do it. And that's actually helped turbocharge and identify a lot of the various gaps. Now that mindset is trying to be scaled across JP Morgan. It sounds almost like this is your ultimate R&D department in many ways, that they test things out and sort of work out at least what they want, maybe not how to actually sort of do it. Are you seeing things like from that sort of personal AI assistance that suddenly now tools like RPA or things like that that were perhaps being promoted five years ago are kind of null and void now, because people know exactly what they want, exactly what connectors they want, exactly what tools they want. Yeah, well definitely, generative AI certainly pushes the horizons of traditional RPA. The mindset of RPA is which is that look at a discipline process and which part can we automate through rules. That's still a very important, important skill set. But what generative AI has done is now where RPA sort of capped out, because it ultimately couldn't be done through rules or automations required a human, now you can insert an AI agent to bring the intelligence of the cognition to complete that part of the puzzle. And so that's what's now allowing, I think, everyone to take a look at their end-to-end processes with a whole fresh lens. And it taps into a new opportunity set that wasn't available before. Derek, I'm curious on this agent economy. Obviously, it's become the envy of the industry, right? It sounds like all banks are now trying to emulate this. How far are you allowing employees to go with this, right? So you talk, obviously, about in internal governance, they have the permissions that they should have. But are they allowed to go externally and pull in Bloomberg through some sort of MCP, bring information in, where are the guardrails? What are the limits? Yeah, so there's multiple levels of guardrails at the lowest level is the guardrails around the platform itself. And it's been constructed as quite a well-governed platform such that what can be done in the platform is safe within the platform. So we don't allow, for example, anyone to come and just call an MCP server. A MCP server would need to be onboarded into the platform and made available. And the prerequisites for that would be security testing, making sure that legal agreements are in use because third-party data also has legal provisions around how data can be used with generative AI. So all of those things need to be checked. But then once it gets onboarded, that's a very powerful step. Because now, centrally, that onboarding has basically onboarded a new capability for the whole firm at scale. And then provided that people are entitled to it. And not all services or data solutions have firm-wide data subscriptions. Some of them do, some of them don't. But again, whatever his employee is entitled to, can then be used in a variety of different ways. So let's move into adoption, Derek, right? So again, you've reached an incredible scale with, I think, roughly one and two employees. So 200,000 or so of 400,000 using these tools daily. But you mentioned initially getting to about 30% with your early adopters. So what did the full court press, from CEO town halls to social proof? I know you worked on this stuff, you're working on this stuff yourself. Look like to get over the 50% mark. - Yeah, so we saw an interesting human phenomenon as we were initially rolling it out. So when new groups of users were being given access, what we noticed was that it didn't matter what team you were, whether you were sales, finance, technology, operations, it didn't matter. About 25 to 30% of people just got busy using it right away. At that point, I sort of, I reflected and I said, "Ah, that must be your early adopter population that you see in the technology." So getting to sort of 25 to 30% of the firm using it actually came quite easily because they just, just used it once they had available. What we, the journey then from 25 to 30% to where we are now, which I think is a little north of 60% or so, has been a consistent journey almost week on week. We continue to see growing levels of adoption of use, but it's been a different slope in that growth. And I think what's going on here is that this is really your, sort of your fast follower population. And so they're sort of seeing their neighbors using it, they're hearing stories about it, also the capabilities and the solutions available are becoming more powerful. And so they're following and using it. We are still on this trajectory or a week after week, this grows. And I do think over time we will eventually be seen 100% of the firm use it. It did take a full court press to make sure that people were and are continuously aware of the new types of capabilities that are available. When we first launched, our HR team was a terrific partner, and they helped build for us what was also, I think, a fairly advanced and pioneering training program, designed for the employee base at large with the objective of making them aware on how to make the best of their AI tools that were available. It was branded AI made easy, and we had tens of thousands of people take that in the first few months. And that was the basics of just what AI can do for you, what it can't do for you, how to use it in basic ways, but all the way up to even more sophisticated advanced concepts like role-playing where you can use it to create, but then you can pivot a personality to then to critique it. Those are fairly advanced methods. And then we've kept that training alive so that as capabilities grow, we just continue to stack on. So people are always being made aware of how to make the most of these things. But then there's a week we tried within JPMorgan to just use every angle to get to the employee base. So CEOs would talk about it in town halls. I personally spent a lot of time in senior leader off sites and things like that, giving demos, talking about what it could do. We have a solutions team that engages teams to educate and ideate. You could show up and you'd see on the screens in many of our headquarters offices, campaigns of what could be done. Our JPMorgan internal newsletter would be featuring stories, user stories. We would, for your user stories, about great stories on what people have done with it and how they're using about it. 'Cause we found that type of peer insight is what really, really captivates the attention of people. So I think all of these things have been very effective in helping drive forward the awareness and innovation. - Do you think as you go along and you get to sort of maybe the last 30% last 20% of people, what's gonna convince those people, right? Like it sounds like you've done a very much a top-down kind of full-code press. But is it going to be the people seeing the person next to them and realizing, oh, this is not intimidating. This is something that can actually help me. I'm curious, like, how do you deal with the lardites, right, of people who are resistant to this? And there's definitely a section of the population. It seems in every organization that is resistant to this. - But I actually think that we'll get to 100%. And the fundamental reason is, first of all, JPMorgan has a culture of sort of wanting to be the best and wanting to be on the frontier of everything. And so when, and if you give people good tools that just obviously make their life easier, and better, I think people are naturally going to be attracted to that. I think that sort of last 20% you're recording are probably people who are still a little bit nervous about it. They don't necessarily know how to do it. Maybe they feel they're going to fail trying. But I think that is there colleagues around them who can use it, it becomes more and more than norm. I think you'll just see that everybody begins to embrace it. I would also add as well that from where I'm focusing, it's not necessarily just the number of users. It's interesting if you take a look at the usage across the population. It's also worth taking a look at those initial 25% of the people and see what's become of them. And it turns out that they're basically become the power users. So they're not just users, but they're using it multiples more than the average. And that's also an interesting population to pay attention to and tap into, because they're the hopeful solution developers that we want to tap into, that they think and build great relevant solutions for the rest of the population. Do you see that being any particular groups, like a client advises all over this, 'cause it makes their job easier and perhaps back off as people less, so something like that is they, are they patents? No, it's all over, it's all over. And they tend to be individuals who are just a little more comfortable with innovation, a little more comfortable trying things, a little more wanting to push the frontier of things. But we are identifying that and that in itself has become a community that we've tapped into. So for example, when we're trying out new beta features and releases, we tap into those communities to give us feedback. And then we want to give that community tools as well that they can use to share their innovations. Derek, when we're talking with other enterprise companies, they've hit a wall, right? That there is this kind of known wall in the industry. You hit this 25 to 30% wall. It sounds like you kind of hit that first wall. Maybe you don't call it a wall because you continue to see that growth, but it was a different slope. Is there anything else that your tech builders, right? You're the builders of these AI agents are doing to help this adoption. Is there something around the interface, other things that you've seen working? Or is this just cultural, beat them with a stick over the head to get them using this? We are here to solve business problems, not here to build great tech. And importantly, the way that that manifested in terms of our product development cycles is that we don't start a new feature, a new product unless we have line of sight to five different teams who know that that capability will solve a problem. And that turns out to be a really important north star for everything because then as we're designing the experience, as we're building the underlying tech components, as we're testing it, we're always making sure that it's solving the problem. But we also make sure that it's not solving only that problem. It's being solved in a general enough way that it can solve other problems, but at least then we know that when we launch something new, it will solve those problems. And I think that that's been really very helpful because then when we also launch a new capability, it doesn't just become a capability in isolation, we immediately attach it going back to some of the change management and communications and campaigns and new stories. We immediately can then say, here's how some of the new capabilities are actually being used in the businesses to solve problems. - I was struck by your personal anecdote that you shared with us earlier before sending an email you pause. You know, any email to say ask what's the vacation policy at JP Morgan or how much vacation do I have or name your other process question. You say what's missing from my AI ecosystem, my knowledge base that forces me to do this. And so how are you trying to inculcate that sort of mindset, where you're actually pausing before you're asking your colleague a question and actually seeing if you can get it from the AI system first. - Yeah, I developed that force myself to develop that mindset about a year and a half to two years ago. By the way, the reason that I sort of did that was it was a reflection 25 years ago when I first was exposed to Google. And I remember at that time, I said, ah, I said, you know, today people are satisfied by just being in a bar and asking a question and not knowing the answer to it. I said, at some point that's probably going to change where that's going to become not tolerable. Everyone will immediately go to the internet and ask the question. And what I observed was that actually played out, but it took, I think, about two or three years for that sort of cultural transformation. And so I witnessed the pattern here again, and I said, well, I don't want this to be two or three years. Let's just do it right now. So that was when I set the mindset. But then what I did was I promoted that mindset. So when I do my senior leader meetings and town halls, first of all, I share the anecdote. But then one of the last departing thoughts that I leave with a group is, what would your ideal, AI assistant do for you? Because that's the mindset that you have to have. If someone can't articulate that, then it's very difficult for AI to help them. They've got to be able to articulate, what would AI, what would they like AI to do for that? Then you can have a structured conversation once that is articulated, what's the data, what's the techniques, what's the connections, et cetera, and begin to unpack it. And so that's the very, very first step of tapping into this flywheel. So Derek, so you've got this Asian economy. You're clearly has some momentum. I'm kind of curious what's next. Where do you take this budget? Where's that budget going? Can you fill us in? Within JP Morgan, we talk strategically about a top-down strategy and a bottom-up strategy. And the bottom-up strategy is really everything that we've talked about today. So building incredibly powerful platform capabilities that have this connections and have the knowledge and have the capabilities that will just grow and scale over time, enable and enable more and more and more. That trajectory, I think, is quite clear how that will play out over the next year and two years. And it will just allow us to do more and more and more, especially as the models just continue to get even more powerful. But I think what's also now been realized is that this bottom-up innovation is great as it is and exciting as it is, won't actually fully transform a company on its own. Businesses run on long processes that cross multiple different types of teams. And so if we want to be able to really move the needle on those processes, think reducing the end-to-end time to disperse credit, or reducing the end-to-end time to onboard a new employee, these types of things. There also has to be a little more of a strategic element to actually rethink what the process itself will need to look like in a world of AI and in a world of AI agents. And so that's a piece now where we're beginning to also revisit with a lot more in intensity and it complements what's going on bottom-up. So as we can begin to rethink these end-to-end processes on what's possible, leveraging many of the componentry that's available, that's, I think, how we achieve Morgan think we'll be able to really, really transform the enterprise. >> Trying to question, what's the one thing that you want to do with AI that you can't do yet? That's a good question. >> The things that I'd really love to be able to do would be to throw vast amounts of documentation, think about like an investment banking deal or something like that, and really have AI be able to analyze that and just eat all of that content insight. And we can certainly do a great job on that with powerful and gestural pipelines and building specialized agents to do things. But typically speaking, when we try new things, we're still sort of running up into some of just the capability limits of today's current models. But boys, they're getting better. You know, what you can do now with the latest models and their very long context window and multi-modality are things that we couldn't do six months or a year ago. Having a system that can just use that with the superhuman judgment would probably be a pretty big unlock. Derek, it's been wonderful having you on the show. Really interesting to hear what you're doing at JP Morgan. Sounds like you've got a, you know, you're very well along the road, but it sounds like you've still got plenty more things coming. So we look forward to talking to you again in the future. >> Thank you very much. Appreciate it. >> Thank you, Derek. >> And special thanks to our presenting sponsor, OutShift, by Cisco. You can learn more about their work on the Internet of Agents via their Linux Foundation project, agntcy.org. This open source project enables agents to work at scale across any vendor or framework with trusted and secure discovery, identity access and observability. For more stories about the AI revolution, like and subscribe to the podcast and check out VentureBeat.com to sign up for our newsletters. [MUSIC] (upbeat music)