From AI Assistants to Autonomous Enterprises: What It Takes to Make Agentic AI Real with Dell Tech

Dell started with 900 internal AI projects. It canceled all but 13 of them.

That may be one of the clearest lessons yet for enterprises rushing into the agentic era: scaling AI starts with knowing what not to build.

At the Six Five Summit 2026, David Nicholson opens the AI Infrastructure track with John Roese, Global Chief Technology Officer and Chief AI Officer at Dell Technologies.

Roese draws a sharp distinction between generative AI’s first wave and the agentic phase now underway. Chatbots made proprietary data easier for people to access. Agents go further by moving entire categories of work from humans to machines.

His framework divides every job into five types of work: productivity, hygiene, coordination, expert, and human element.

As agents absorb some of those categories, the job itself shifts toward what remains. Dell’s engineering organization offers an early example. Coding assistants initially automated productivity tasks such as writing code comments. Agentic tools are now taking on hygiene and coordination work, leaving engineers increasingly focused on defining specifications, applying expertise, and exercising human judgment.

Key Insights:

🔹 Roese separates genuine agentic deployment from “agent washing.” Chatbots and digital assistants help humans complete work; autonomous agents take defined work off a human’s plate entirely.

🔹 Agents do not simply “take jobs.” They absorb specific categories of work, changing which skills become most valuable within each role.

🔹 Dell cut 900 internal AI initiatives down to just 13, treating disciplined prioritization as a prerequisite for deployment at scale.

🔹 The company uses four to five distinct sources of AI tokens, spanning on-premises open models, frontier APIs, and on-device personal-agent frameworks. Each workload is matched to the right combination of cost, performance, privacy, and compliance.

🔹 Roese expects roughly 70% of Dell’s agents to be “headless,” operating without a dedicated human owner. That required a new security model because traditional identity systems assume every actor ultimately sits beneath a person.

Dell now gives every autonomous agent its own company-issued digital identity, fine-grained permissions, and an immediate kill switch.

In an agentic enterprise, identity can no longer belong only to people. Machines need to be governed like members of the workforce, too.

Watch the full video at sixfivemedia.com, and subscribe to our YouTube channel so you never miss an episode.

Explore more sessions from Six Five Summit: AI Unleashed 2026 at sixfivemedia.com/summit.

Disclaimer: Six Five Media is for information and entertainment purposes only. Over the course of this video, we may discuss companies that are publicly traded, and we may reference their equity share prices. Nothing discussed during this webcast should be considered investment advice or a recommendation to buy or sell any security. We are not investment advisors, and you should not rely on this content as financial advice. Six Five Media collaborates with technology companies and industry leaders to produce research-driven interviews and multimedia programming for enterprise technology.

‍

John Roese:
What we've realized is the first principle of token economics in the AI agentic era is you have got to have a diverse source of intelligence. You need to have sources that range from different compliance levels, different capabilities, but also different economic strata.

David Nicholson: 

Hi everyone, and welcome to the Six Five Summit AI Unleashed 2026. For this AI Infrastructure Track Opener, we're exploring how enterprises move from AI experimentation to operational transformation. Joining me is John Rose, Global Chief Technology Officer and Chief AI Officer at Dell Technologies. John, welcome to Six Five. Great to be here. There's been no shortage of excitement around AI and a lot of organizations are talking about what they're going to do, moving beyond pilots and proof of concepts. What convinces you that agentic AI is really ready for production?

John Roese: 

Well, the big thing is that I'm actually doing it, so I feel comfortable that I actually have agents in production at the company. But the bigger thing is just like we are seeing that pattern hole. We are seeing the technology mature. At the same time, just to be really clear, the amount of agent washing going on right now is staggering. I mean, if you ask a random person what an agent is, I don't know what they answer, but it's probably wrong. We are confusing chatbots and digital assistants and very traditional approaches to just unlocking data with autonomous agents. And so it is incredibly important that people realize that these are two different things. And the most important thing to recognize is we graduated from an era where almost all of the generative AI work in enterprises was about unlocking proprietary data with generative capabilities using chatbots. That's great. You should do that. That's super important. We've had a tremendous impact on Dell by just unlocking our proprietary data. Completely decoupled revenue growth from cost structure. Fantastic. Worth doing. But the second phase, where we move into Agentic, is different. You're not just unlocking data. What you are doing is digitizing work. You are literally shifting work from a human being to a machine. That work may be very simple. It may be just an autonomous task. Book my travel. Summarize this thing. But it's done without any kind of human guidance or oversight. Or it could be much more significant. Clean up my CRM data. Build this software for me. Those are very, very different worlds, and I believe that one of the challenges is people have not quite figured out that that breakpoint was pretty abrupt. It's different infrastructure, it's different technology stacks, it's actually a different objective. One, unlock proprietary data to make humans more productive. The other… decouple human capacity from work capacity. That's what's going on at Engentic. Now, the reason I'm confident that exists is we are now, we built our first agents two years ago. We put them into production over the last year. We feel like there are enough examples of actually using these tools in very carefully targeted, specific ways, following a good governance process, that they actually have a material impact. I have agents cleaning up CRM data. I have agents writing code. I have agents doing all kinds of, I have agents doing special pricing. I found that if you find the right process and the right work and you use the technology correctly, these do get into production. And when they get into production, the biggest impact is in that first phase, you could expect 20, 30, 40% productivity improvement around a task. In the second phase, when you start talking about agentic, it's orders of magnitude changes in how fast and how effectively you do the work. And that is the exciting part. We're very early. We don't expect everybody in the world to be exactly at the same place, but it's a different era. There's demonstrable technology in place. If we sift through all the noise and are very precise about the definitions, it really does create an enormous opportunity for everybody.

David Nicholson: 

Along with that opportunity is, you know, concerns when people read headlines, because often the terms task, job, work, employment, all sort of get conflated with one another. When you talk about the things that agents can do that humans might have done in the past, I want to hear your perspective on that. Are we talking about going in and doing the stereotypical decimation? One out of every ten humans is no longer necessary?

John Roese: 

Yeah, we didn't do ourselves a lot of favors as an industry with just kind of hyperbole, but there's some reality underneath this. Let me try to cut through the noise. When you think about agents, they are a piece of technology that is capable of doing work, but I use that word very carefully. I didn't say a task. I didn't say a job. I said work. What I mean by that is agents are good at doing the effort to achieve an outcome in a certain category. They are designed to do work of a certain type. They might be very good at hygiene, low autonomy or high autonomy, low complexity, rinse and repeat work. They might be very good at coordination work. They're trained to do that. They might be good at productivity tasks, but the bottom line is they are a piece of technology that does work. By the way, there are other technologies that we've used before that do work for us, a word processor, any kind of automation tool, internal combustion engine. We've been on this journey of shifting work away from human beings and into the machine layer for a very, very long time. And AI agents are just another step in that. The reason that word is so important, though, is that we have to basically put it in the context of those other words you used. For instance, human beings do work, but they do work in the context of a job. A job is a container of work that the person does in their day. Inside of that container, there are multiple types of work. In fact, we've done studies and realized there are five kinds of work that seem to happen in aggregate. There is productivity work, the act of doing kind of a simple repeatable task, hygiene work, the act of doing a long running kind of below the noise floor task, you know, clean something up, do this thing, write code as a hygiene task. There's coordination work, make sure this process happens, drive this to completion. There is expert work, do something that requires a specialized skill to achieve an outcome. And there's something called human element work, which is all the stuff people do, interact with this person, influence this person, sell to this person. You take those five categories of work, independent of agents, and look at any job in the world and take it apart and ask how much of that job are those five categories, you will find some very interesting things. You will find that every job in the world has probably at least three different types of work inside of it. There is no job that's just productivity or just hygiene because you're talking to people and you're doing other things. There are many jobs that do not require any expert level or skills. That's not required. But there are many jobs that require deep human element work to be successful. The reality is when we think about what agents are going to do to the workforce, they don't take your job. They change your job. And they do that because now they can do some of the work that is within the job as a machine. So they've removed that work from your container called a job, and they've put it below the noise floor, and they've done it for you. And the effect that has on your job, if suddenly your job included all these different kinds of work, And then afterwards, it only included this much work because everything else was being done for you. What happens to the job? It doesn't go away. It evolves. It changes. You concentrate on the expert and human element work, and you stop doing productivity, hygiene, and coordination work is a very good example of that. That shift is incredibly important to understand because if you keep it at the job level and think agents are going to take your job, first, you're factually incorrect. There are no agents in the world that can do all of the work in a job simultaneously. They just don't do that. There are many, many agents that can do pieces of the job, meaning certain types of work, and when they do that, that work drops below the noise floor, and the jobs that remain, the work that still is within the job, becomes your focus. Let me give you a real tangible example of how this has already happened. If you look at software development before Agentic and after Agentic, before Agentic, we were using coding assistants. They weren't agents. And what we did when we applied an early coding assistance to our engineering organization, if you think about those five types of work, Turns out, the only thing they really affected was the productivity piece. They took a bunch of productivity pieces of being an engineer, code annotation, comments, and they made those go away. And it was worth doing. It took a lot of time away from the engineers, wasting it on this boring stuff, so they could focus on coding. That's what we said. They were going to focus on coding and not do the other stuff. So they weren't doing productivity work, but they were absolutely doing hygiene work, which is coding. They were doing coordination work. CICD was still a manual process. They were doing expert work, developing architecture. They weren't doing a lot of human element work, because they're kind of introverted people. Fast forward into today, where we have spectrum and development agentic coding. What has changed? Well, the productivity work is still gone. It's actually more gone. The hygiene work is gone. Hygiene is the act of coding. The reality is the agents are doing that for you. The coordination work is gone. The CICD pipeline is run by an agentic workflow, not you anymore. And so the only thing left for you is you are now the spec driven development architect and your job is now to have expertise to build the spec. The funny thing that happened is you actually have to pick up human element skills because it turns out to write a spec, it's not just the technical parts that are in the spec, it's product requirements, it's market requirements, it's interaction with customers. So the very definition of an engineering architecture before an architect before and after was in the old world, they did kind of all of this stuff partially and they're mostly a technical expert. And in the new world, they did none of this stuff and they were mostly an expert that had good humans interaction skills and could interpret requirements and turn them into a spec. Are those the same job? They are not. They are the same person that has evolved, and they evolved because agents had materialized and removed part of the work from the job and forced the job to evolve. Last thing I'll leave you with, though, is if you believe what I just said, which is exactly what we're doing at Dell, exactly what seems to be playing out, one thing it really does is debunk this idea of anthropomorphizing agents, this idea that agents are part of your workforce, that agents are digital humans. They are not. In fact, it makes about as much sense to put an agent as I described it in your org chart as it would have to put a word processor in your org chart in the past. And so that is a very different way of thinking about it, but we think that is actually what's happening. And so because we're getting to the point that these are becoming real and we're mature about how we think about them, I, for the first time in three years, actually have a reasonable view of what the workforce at Dell looks like, what the jobs of the future are, and what I will tell you is every job is going to change because all jobs have work that agents can extract and do, but the jobs that remain, and there will be lots of them, are an evolution of the jobs that exist today accommodating for the fact that a whole bunch of the kind of boring stuff and the things that aren't necessary to be done anymore disappear below the noise floor and humans shift towards the high value work. And I actually feel like I can see through the fog for the first time because we've done this work. By the way, we did this by analyzing 6,800 jobs, by basically using agents to analyze this data, by getting a lot of empirical data and then by going and doing it. So this isn't a theory, this is actually what we're doing. me that the nuance around agentic, specifically vis-a-vis things like jobs and work, is incredibly important.

David Nicholson: 

Yeah, I have to say that's the best explanation of, you know, from an individual perspective, the direction that AI, specifically agents, is actually going. Now, beyond that, you sort of touched on this a little bit, Dell as customer zero, Dell figuring out for itself how to deploy these technologies effectively so that you can then work with clients to help them develop. What about this idea of organizational change? Because, great, you've just laid out an amazing explanation for how agentic workflows will change our individual relationships with work. But what about the organizational change that needs to take place? What is Dell seeing and how is Dell helping customers on that journey?

John Roese: 

The entire AI journey for the last, let's say, two and a half years has required significant organizational and cultural change in every company, and we are no different. It started with, you know, the first thing we learned a couple of years ago. You've got to be able to You've got to have governance. You've got to do this top down. You cannot do this as a suggestion box. You cannot have random chaos and hope you get to an AI outcome. It's just not possible. You have to pick your battles. You have to target them. And because of that, you have to understand your business process. You have to understand where your value is. You have to be very precise. And then you have to realize that you're not going to be able to serve everybody. I had 900 projects when I started. We canceled them all and did about 13. And those 13 turned into decoupling revenue growth from cost structure and a significant impact to the company. That was a cultural and organizational change. Top-down culture and organization versus bottom-up culture and organization. So now we shift into the agentic era, and now we're proliferating this technology directly into jobs. And all those things were kind of tools. Now we're changing where work is done, and that will require significant change. But the biggest changes that are going to happen there are the ability for people to understand that the effect of these autonomous systems is not incrementilization. It is rethinking of the entire organization and the work. That's why, by the way, you use the word task. I don't use that word to describe agents. If you narrow the agent to do a task as you define it today, you're underestimating what it can do. You're limiting it in a way that is actually counterproductive. If you want to do RPA, go find a task and apply a script to it. But if you want to do agents, you go find work. you find higher level abstractions that are outcome driven and you apply agents to do that work. Now, think about what culturally and organizationally that requires. It requires you to be extremely open-minded about the fact that every job is going to change, how you do work is going to change, who does that work is going to change. That is a significant effort to go through that exercise. It is not easy to do. That's one of the reasons why we did not go and randomly throw agents all over the company. We've been very precise about targeting them in places where we can work through these changes with the organizations. We're now at the point that we are scaling that, but we are doing it, based on the description I just gave you, with a repeatable framework. We have clarity about what this means. It's understandable because you're talking about fundamentally changing the structures of how a company is organized and what people do within that company. Don't take that lightly, but realize that it is going to change. If you proactively change it, what ends up at the other end is a highly productive company with people doing work that exists on top of an agentic and AI foundation that is much more efficient and powerful than you've ever had. and usually an environment that is much more fulfilling, much more productive, and much more effective and successful. That's a great goal to go after. You don't get there by incrementalizing on the past. You have to be willing to rethink every process, be open-minded to changing every job, and changing the organizations around it. That's a big deal. We've never had a technology that did this to us, maybe since like the Industrial Revolution. We have to be in a mindset that we're willing to do that, but again, good governance, targeting it, working on trying to figure it out before you go broad, but don't take two years figuring it out. All of these things are learnings that we've had. And the result is when you do them, again, you decouple human capacity from work capacity in your company. And that causes just explosive growth to occur. Orders of magnitude improvement in productivity in the dimensions where you apply it.

David Nicholson: 

Let's talk about the investment in all of this moving forward. There's a saying in financial circles that if you borrow $10,000 from a bank, you're a customer. If you borrow $10 billion from a bank, you're a partner. Increasingly, just the raw number of agents that are going to exist in an organization make this more of a partnership than a customer-vendor relationship. When people are considering ROI, it is the investment. How do you quantify that investment? We start talking about things like tokens, the generation and consumption of tokens. How should we think about that, about the investment in AI?

John Roese: 

Yeah, let's be really clear. Nothing is free in the world. AI is no exception to that. AI is powered by compute, it's powered by data, it's powered by processing, and that processing has to exist somewhere. There's a cost to it. That, hopefully everybody gets that. But when you think about applying AI now to the broad and diverse work that makes a company run, you have to have a diverse approach to this. You have to have choices. You cannot do that with a monoculture. Because think about it. If I describe five different kinds of work that happen in a company, and then I ascribe economic value to them. So let's compare the two. If I build an agent, and that agent lets the CEO of the company make better decisions in real time to guide the company into the future, how valuable is that agent? Pretty high. If it costs me $100,000 a month to run that agent, it's probably worth doing. On the other hand, if I build an agent and its job is to clean up CRM records, and each CRM record it cleans up has an economic value of $0.50, then while there's millions of them, I'd better have an environment in which it is cost effective to do that and it doesn't cost $12 in tokens to do a $0.50 task. That's the spectrum that we're dealing with here. And so what we've realized is the first principle of token economics in the AI agentic era is you have got to have a diverse source of intelligence. You need to have sources that range from different compliance levels, different capabilities, but also different economic strata. At Dell, I have four moving to five different fundamental sources of tokens. They have very different economics. I run open models on-prem in my data centers. I run frontier models in my data centers. I run frontier models in VPCs I control. I use APIs. I now run models on devices using things like personal agent frameworks. Each of those are not arbitrary. When you look at them from either their economics, their performance, their regulatory and compliance risk, and their functionality, they are different. Now you would say, oh, this is really complex. It's not. It's complexity by design. Because when I look at a piece of work, comparing that CRM agent to the agent that's powering the CEO, I now have a choice. I don't have just one answer. Imagine if you were the customer that wanted to do those two things, but you had adopted a single provider with a single set of models over a single economic model. you fundamentally wouldn't be able to do the CRM agent because it's just not affordable in that model. And so we've come full circle back to the hybrid architectures of the world, which basically say hybrid is not arbitrary, hybrid is choice, hybrid is diversity. And when you're applying AI technology to jobs and work, they're affecting the jobs and they're changing the work. Is your work and your job structure homogenous? Is every job equal? Is all work the same? Of course it isn't. Is it done in the same place? So having an infrastructure and an AI capability that gives you the ability to map to the right economics, the right control, the right framework, the right compliance, these are all things that are absolutely essential. For us, I sound like a broken record. You've known me a very long time. I can show you presentations from almost two decades ago where we said hybrid is the only answer. It makes no sense to think of whether it was the cloud or a one cloud to rule them all or a single IT infrastructure. And AI has just amplified that. I cannot imagine trying to execute AI in a monoculture because what you're executing against is a highly diverse set of activities that make a company a company.

David Nicholson: 

Yeah, no, it makes perfect sense. And as I sit here as sort of a proxy for the CIOs and CTOs that I work with, you've taken me through this journey where, okay, we've taken care of this idea that no, no, no, we're not firing all the humans. We've addressed organizational concerns. Now, cost concerns, the economic model, the kind of hybridization, the idea of there isn't one fit for all functions. The thing that keeps me up at night as the proxy CIO is this idea that, OK, I need to deploy infrastructure today that will support all of these things. But I'm being told that we are on the cusp of the post-quantum era from a security perspective. So how do I make sure that I'm living up to my fiduciary responsibility from a security perspective today, but also setting myself up for success in the future.

John Roese: 

Yeah, I mean, there's two pieces of that. Let me take the post-quantum one. That's the easy one. Look, we have done a good job as an industry to kind of develop post-quantum algorithms. They'll be very precise. You know, quantum risk from a cyber perspective is very specific. It is the ability to factor prime numbers in a quantum system with the right algorithms and right scale becomes very easy. The foundation of asymmetric key management protocols in cryptography are based on that. If that math becomes easy, those things break. RSA breaks, a bunch of other stuff breaks. Four years ago, we started building algorithms. The algorithms are available. They're starting to roll out. We still have a reasonable amount of time. There are some issues around capture now, harvest later. So it's a very real thing, but you shouldn't panic over that. There is a fairly good top-down industry-wide effort to bring improved key management protocols into the places that matter. If your data is living in your data center and doesn't ever egress, you don't have a problem right now. If your data is flowing into a public cloud across a public interface and you're using very weak encryption protocols and key management, you probably ought to fix that because that's a target. And so the reality of it is that one's easy. The cyber discussion in general when you move into the AI era is a bigger problem because Just like agents are a different technology that does work in a different way and exist in your environment in a different outcome, guess what? They have a different security model. They are not exactly the same. And in fact, what we've learned, and this is one of our customer zero experiences, early on, beginning of last year, we started kind of bringing the industry together to say, we don't have any idea how to make agents talk to each other, how to make them secure. Our assumption that they're magically secure is a poor one. And so we spent the last year and a half working with our security partners and with the industry to try to figure out ways to standardize protocols. We now have protocols like A2A and MCP. Those are, in general, questionable in terms of how secure they are, but they're at least standard. So we can now secure them. The biggest learning for us, though, is we realized that if agents do work, and that work is done on behalf of people or organizations, then it is essential that we be able to control them. And so one of the decisions Dell made last October, and we now enforce as part of our agenda guidelines, is that all agents that are running autonomously, touching our data, whether they're internal or external, carry a Dell-issued digital identity. We give them the identity. You have permission to touch my data, to interact with me based on an identity I grant you. Now, the advantage of doing that is that if you go berserk, And I revoke your identity because tied to that identity is fine-grained access control, authorization. I can make you disappear even though you're running on a third-party platform or you're outside of my environment. We made that decision a long time ago. Now that decision is starting to become fairly common because it gives you the kill switch. In fact, in the EU-AI Act, there was a requirement for kill switches. If you ever wondered how to do it, I'll give you the easy answer. have a universal control over agentic identity that is consistent with your overall identity management framework, and if you control the identity of agents anywhere that they work on your behalf, you have the ability to introduce a kill switch. In addition to that, we realized there was a lot of fragmentation around telemetry, even the difference between two types of agents. There are agents that are inherited. An agent that works for Dave is one kind of agent. From a security perspective, It uses your credentials, your authorization. It works on your behalf. That's fairly easy to understand. You're accountable for it. We can monitor it as if it's Dave. That's an easy one. But then there are these things called headless agents, which don't work for Dave. Dave might make them come into existence, but then they just go clean up CRM data, or they run a process, or they do something. Those are the ones that are far more dangerous. They probably represent 70% of the agents that will ultimately be deployed. Those headless agents actually were quite hard to incorporate into the security frameworks. because our security frameworks work on a human hierarchy. Remember my comment about the word processor? Agents aren't necessarily going to be in your org chart, but if the way you ascribe identity and access control is to the organization of the company and the agent isn't in there, you have to come up with a different way. And so we actually worked with partners like Okta and Palo Alto and others and came up with ways to say, how do you deal with headless agents? All of this is now starting to become available and real, but the principle here is, Just like as you move into the post-quantum world, something changed that you had to adapt to. Post-quantum cryptography, as you move into the agentic world, something changed. Agents that now do work independent of humans, and you have to adapt your security architecture for it. The good news, all of this is doable. I think we're early. I think we're ahead of it to some degree. But the one answer that will absolutely fail is to just do nothing. To just assume that these new technologies show up and your existing approach to however you've run your IT organization forever is sufficient. It is not. It requires a movement to a different kind of hybrid architecture, an evolution of your security architecture, a rethinking of your organization, a rethinking of fundamentals of work. You know, nothing going on here. Just, you know, everything's changing. But if you change it But programmatically, if you work through it and you use people like us and others in the industry that have kind of done some of this stuff, it is navigatable. And the result is you disconnect this relationship between human capacity and actual work and the work capacity of the world. And the effect are things like happened to Dell, where suddenly your revenue is doing this and your cost structure is doing that at the same time, which has never happened in the history of business. And yet now it starts to become something that's accessible to a lot of the industrial world.

David Nicholson: 

Fantastic. Thanks so much, John, for joining us. 

John Roese: 

Thanks. Good to be here. 

David Nicholson: 

This has been our AI infrastructure track opener. To our viewers, don't forget to hit subscribe, follow us on social media, and check out all of our Six Five Summit content at sixfivemedia.com/summit. See you soon.