Agentic AI in the Real World: ROI, Autonomous Agents, and Tokenomics

Courtney Hughes (00:00.814)
Hello and welcome back to the Dell Technologies Forum Podcast series where ideas and innovation meet acceleration. In this series, we bring together Dell's brightest minds and industry pioneers to help you navigate the ever-evolving landscape of technology. I'm your host, Courtney Hughes. And today we're kicking off our fourth season with a conversation that I feel is very timely, strategic, and truly grounded in what customers are wrestling with right now.

Moving from AI curiosity to real business outcomes, preparing for autonomous agents, and understanding how infrastructure strategy shapes speed, control, and cost. So I'm thrilled to welcome back John Roese, Global CTO and Chief AI Officer at Dell Technologies to our podcast. Welcome back, John.

John Roese (00:52.684)
Well, great to be here again.

Courtney Hughes (00:54.314)
It's great to have you again. John, we met last year before Dell Technologies Forums and AI adoption was on everyone's mind, right? What's your take on how the industry is doing on narrowing down the use cases, adopting AI, and seeing real ROI from those efforts?

John Roese (01:12.684)
Yeah, I mean, I'm seeing progress. We're not there yet in every enterprise and every part of the organization. And you know, there's still a distribution curve of people who are making great progress and people who haven't started. But year over year, more and more of the enterprises of the world are moving out of experimentation. But maybe they listen to us, interestingly enough. We were very clear having a few high priority projects tied to clear business processes with clear ROI is far more actionable and impactful than trying to do a thousand AI projects randomly and hope you get it right. I think the correlation has been companies that have strong governance, that have focused on prioritization, really leaned into manageable and measurable ROI, they are tipping into production. And I'm very encouraged that, you know, probably, I don't know, a third of the conversations I have, customers are now demonstrating that this is starting to move the needle for them and

You know, we know what happens when that occurs. It is transformational and I'm  optimistic and feeling like we're  starting to see that momentum materialize in a way that was very different than a year ago.

Courtney Hughes (02:12.184)
So is it true to say that, you know, the winners are not the ones doing the most AI, but the ones that are being the most intentional about where they apply it?

John Roese  (02:20.93)
Yes, without a doubt. You know, my word of the year this year, last year it was agentic, this year it's governance, because the correlation is if you have governance, which is not a way to prevent you from doing work, but having a clear way of prioritizing, of choosing the most important things and then focusing your energy on getting them into production, that correlates directly with the people who are having success. It's very, very clear that that's the path to production if you want to get this right.

Courtney Hughes (02:46.252)
Yeah. And so with those customers that you're talking with, John, how does Dell help them move faster with a business first focus?

John Roese  (02:54.432)
You know, on a couple of different dimensions. You know, we leaned in and sharing our journey of helping people understand what we were learning. And I think, you know, the first dimension is by by just giving people an example of how to get out of the POC purgatory they were in and into production by having strong governance, by having prioritization, has been incredibly helpful. I have many conversations with chief AI officers, boards of directors, CEOs, to talk about that. But the second thing that's happened is the maturation of the tech stack.

You know, AI factories and the underlying technology ecosystem we pulled together means that when you actually figure out what you want to do, it is not a do-it-yourself project. You do not have to build your own server infrastructure and design your own data management framework and quite frankly, even pick your own agentic framework. We have largely validated and proven those out. And so the, you know, the goal of industry and our definite goal with AI Factory is to move the line in which there is customization up higher, so that customers can take advantage of what's happening at scale in the industry and then build on top of it. 

So on one side, we're showing people a path forward that really is about philosophy and sociology and organizational design and prioritization and governance. And on the other side, we are absolutely making it easier to get those ideas, once they define them, quickly into production by basically bringing the tech stack and the ecosystem together in a way that was not done a couple of years ago.

Courtney Hughes (04:18.978)
This year, I mean, we've heard a lot. It's been about autonomous agents. And I know we've been focused on security, identity, and observability. So we know exactly what our agents are doing and communicating. Talk about our federated approach.

John Roese (04:35.51)
Yeah, I think let me be really clear. We are exiting a first phase of AI, which was really transactional AI, that we were using AI tools, chatbots, rag systems, et cetera, into the first phase of autonomy, what I call tactical autonomy, which is really code for: we are now delegating work to things called agents that actually work on our behalf, that there is not a human fully in the loop, and that many much of the work is done without human oversight or intervention.

There will always be human intent and human validation in any autonomous system, but the middle, the part of actually doing the work, is shifting more and more into a level of autonomy that we've never had before. That is incredibly powerful. In fact, to give you an idea, when we just had tactical AI, kind of the transactional stuff in coding, we got maybe 20 or 30 percent improvement of our engineering efficiency by taking a bunch of hygiene tasks off their plate. They didn't annotate code, et cetera. When you move to a agentic coding, which is where we're going right now, you in some cases see a hundred X improvement in the effectiveness and scale of your engineering organization. 

A small team with a set of agents can outperform a gigantic team in a much faster timeframe. That is incredibly powerful, but it is also incredibly different and potentially incredibly risky. And so the challenge is not building the agents, even though that is a piece of it. It's really how do you make it secure? How do you make it controllable? How do you make it observable? 

And so our approach was to start with this assumption that agents are going to be where the work happens. Therefore, you're going to need multiple agentic platforms. They are going to need to work as a system. And so we've done things like define how we have universal identity and universal policy and how we can do observability, how we can have hybrid agentic environments where some of your agents run external and some run internal, yet all of them are under your control. These are new architectures that you know we've largely worked through and are learning a tremendous amount about. 

What we've concluded from that is: It is inevitable that every company in the world moves into agentic because the value is incredibly high. But it is also incredibly important to realize that this is a fundamentally different technology, which will require an architectural evolution on every layer of your stack from token production to infrastructure to security to observability to data management with context layers. And so it's a big deal. But the good news is we now have agents running and they are doing incredibly interesting things. And it does tell me that the actual productivity burst coming from AI, we have not even begun to tap into it and it is going to be even more transformational than you can imagine once you start to put all the pieces together. But it will require work. It will require you to think differently to actually prosecute it properly.

Courtney Hughes  (07:11.682)
You know, the real unlock for agents is trusted autonomy.

John Roese (07:17.332)
Absolutely. Think about this. You know, I don't like to anthropomorphize agents because they're not people. They don't do jobs. They do work. But the reality of it is, is if you had a human being doing work on behalf of your company, maybe very sensitive work, and you didn't even know who they were. You didn't know where they were. You didn't know what they were doing. You didn't care where they did it. You didn't care how they who else saw your data.

You would never go down that path. But yet, because people are in many cases underestimating how significant the changes as you move into agentic, we fundamentally are accidentally giving up the trust that's necessary. Our philosophy, and we've implemented this inside of Dell, is we will not do that. We believe we can move incredibly fast, but we also believe that we can do it in a way that keeps us in control. And my message to every customer is: agents are incredibly important. You should not not participate, but you should also realize that you're gonna spend a great deal of time and energy making sure that you can trust them because if you can't, it is like delegating work to a random stranger on the street and hoping that they don't mess up your company. So we don't really want that to happen.

Courtney Hughes (08:23.424)
One thing I would love your opinion on too, John, because this was a very big buzzword. you know, tokens have been a topic for a while, right? And just three weeks before Dell Technologies World, tokenomics became everyone's buzzword. I would love to know and understand how are we managing token placement and what's important for customers to hear.

John Roese (08:45.346)
Yeah, I think it's funny. I'm an engineer and I usually think my day job is about technology, but in the last six months, probably a good portion of my life has been spent managing economics. We call that tokenomics now in the sense that it's really: how do you efficiently deliver primarily agentic outcomes? First generation AI tools weren't all that token intensive. Agents are. Reasoning models and advanced agents, multi-agent systems -- We have examples where you know, a single agent might use 30 million tokens in a day.

That's a tremendously active agent. We have other ones that might use a few hundred, you know, and it  varies dramatically, but it's significantly more than we've ever seen. So because of that, because tokens have cost, what we've realized is that, you know, again, we need to spend a significant amount of time making sure that the tokens we produce and use and ultimately consume are the most efficient path to actually getting to our outcome. Let me give you an example. 

Today, most people think a token is a token, but the reality is there's dramatic price difference between producing a token in a data center using open models versus producing a token through some random API to a third-party tool that you don't even have any control or contract around. Those two might be three orders of magnitude different cost. And yet the tokens are very similar. So the reality is as we look forward and what we're doing in Dell is, you know, we've created diversity. We believe that, you know, by the way, hybrid is the name of the game and everything in IT in our view.

It's no different in token economics. You need to have multiple sources of supply for tokens that cover different economic strata and ultimately give you different capabilities. On some sides, you may need tokens that run a very specific specialized models, open models in a private environment. That's by far the lowest cost way to produce tokens and get to economic outcome. And you may need that because you may have agents that actually do relatively minor work, like clean up a CRM record and have to do that at a very low cost structure.

On the other end of the spectrum, you may have agents that are developing the spec for a new product, which is the most complex piece of spec driven development. And those you absolutely should use, the most advanced foundation model possible. And because there's things in between those, it's incredibly important that you have a wide range of places where you can consume tokens that give you different capability and different economics, and you have the ability to choose which workloads go where. Today, most people are not doing that.

John Roese (11:07.18)
They are just randomly consuming tokens on whatever tool they have. We believe that modern architecture will ultimately evolve to say it is important to not believe a token is a token. Tokens matter based on what they do and what they need to cost. And the good news with hybrid architectures is we have a lot of choice. I have five different economic models that I can now tap into from a Dell perspective to achieve my agentic outcomes based on token cost and capability.

Three of them are actually fixed cost. There is no per token cost, even though they use foundation models, and two of them have a variable cost around them. That gives me tremendous capability to make sure I'm not the guy that has the $500 million surprise bill because I didn't manage my tokens correctly.

Courtney Hughes (11:53.646)
I mean, it sounds like, you know, tokenomics goes beyond just a finance conversation. It's more of a infrastructure strategy conversation.

John Roese (12:01.238)
Without a doubt, if you treat it as purely a finance discussion, like here's an example. If you go and choose a single provider of tokens with this one AI model, you have absolutely no choice. All you can do is negotiate a better contract. Trust me, you're not going to get a better price. If on the other hand, you have an architecture that gives you multiple sources of supply, then you can do arbitrage between them. Then you can buy great workloads back and forth. You can take a N-1 model and use it because it's more than sufficient and half the price, or you can use an N+1 model because you have the most complex problems to solve. 

The financial equation comes after the architecture's in place because if all you have is one option, it doesn't matter how smart your finance people are. At best they're going to negotiate a slightly lower cost for that one thing for everything. And that will not work in the era of token economics.

Courtney Hughes (12:48.152)
John, you're keynoting several Dell Technologies Forums, Sydney and Tokyo. What do you think the main takeaway for customers will be? And what's the advantage of attending in person?

John Roese  (12:58.732)
Yeah, I think, you know, it's almost group therapy is a good way to describe it at this stage, because everybody knows they need to do this. Everybody knows that it's important, but you've seen the statistics of how many people are actually getting into production. What's interesting at the Forums, because we have a large number of customers that are starting to tip into production, including ourselves, is it actually creates this kind of comfort for people that if you hear about what we're doing, if you hear about some of our bigger customers and our governmental customers that we're working with, and that they're starting to get their first use cases into production, they're figuring out how to navigate many of these challenges and they're seeing the value. 

Now, if you're in an environment that you haven't done any of that, but you're hanging around with a bunch of people who have, one, you're going to learn from them, but two, you're going to feel more confident going back to your own organization and saying, look, there is a path forward. It's going to require some discipline. And I have proof that it works. Those are things that most people didn't have a year or two ago. And what we do find in the Forums, and I have a lot of conversations when I'm there with individual customers, is they're looking for that. They just want to believe that this is possible. And I'll tell you one of the things that happens at the Forums is we prove to you that it is absolutely possible. 

It isn't simple and easy at the level of just doing nothing, but it's also not as hard as you think if you can learn from your peers and learn from us and learn from our ecosystem.
So I just feel like there's a lot of noise in the system. The Forums are a place where it becomes very pragmatic and where we really get into the weeds of how did we do it? How did our customers do it? What are they doing? And that is actually probably the most valuable tool most customers can walk away from because it gives them confidence, hope, and a and a plan to then go on the same journey.

Courtney Hughes (14:37.72)
John, thank you so much for taking the time and speaking with me.

John Roese (14:41.682)
Great to be here. I this is probably the most important conversation and topic that we're gonna have for at least the next day. And you know, by the way, the next time we talk, it'll be another AI evolution, but it's all building on the shoulders of what came before us. Today it's agentic. Tomorrow it will be AI fabrics, which are agents working as a collective. And we don't know what's beyond that, but we do know that AI is gonna redefine our entire world.

And it's incredibly important that people are on the front end of participating in that evolution.

Courtney Hughes (15:13.954)
Yes, I always enjoy talking with you. I always take so many things from every single conversation that I have with you. And I'm sure that our listeners will as well. Thank you all for tuning in. You know, we're in season four, and this season is all about helping customers move from AI momentum to durable enterprise outcomes. 
Don't forget to follow and share the Dell Technologies Forum podcast for more on the technology shaping the way we work. And if today's episode got you thinking, join us at a Dell Technologies Forum near you, where you get to connect with peers, see these innovations live, and take the conversation further. You can find the nearest Forum on our website. Thanks again for listening, and we'll be back soon.