John Roese’s 2026 AI Predictions: AI Governance, Data Readiness, Agentic AI, and AI Factories

So what do you think is going to be the next set of disruptions and changes in the AI world in 2026? 

Welcome to AI Insights with John Roese. You know, today I want to talk to you a little bit about something that we do typically at the end of the year or the beginning of the year, which is try to predict the future.

Last year, we had a whole bunch of predictions — that largely came true, which is good — and, you know, when we think about 2026, there are a whole bunch of other things that are going to happen that we think are important.

Now, first disclaimer: These predictions are not meant to describe everything that’s going to happen. There’s a lot more going on than, you know, the five or six things that we’re going to talk about. But it is interesting to think about some of the things that you may not have thought about that might happen this year, and how they’re going to impact the overall industry of AI.

So let me walk you through them and, you know, love to get your feedback and let’s see how we do this year. 

Prediction #1: A call to action — governance frameworks for a fast-moving ecosystem

Number one is really both a prediction and a call to action. And that is that as we have scaled AI in enterprises like we’ve done at Dell and many other companies, one of the things that is now abundantly clear is that the overall governance of AI in the world writ large — not just in the enterprise but at the regulatory level and the governmental level — is a bit of a mess. And that’s the only way I could describe it. 

Today, as the chief AI officer of Dell, I have a thousand jurisdictions telling me what to do with no coordination between them. We have AI policies that are conflicting with each other. We just don’t have good tools and capabilities, and we definitely don’t have the right frameworks in place, to ultimately make it easy for an enterprise to basically find a problem and execute it without running into potentially third parties or regulatory frameworks that are indecipherable. 

As a prediction, we know that if we don’t do better at governance — both at the global level in terms of political governance, regulatory governance, but also inside of enterprises, being able to choose what you’re doing and making sure you’re doing things that are the right things to do for your company and are solving the right problem and bringing some order to the chaos — AI very quickly spirals out of control and you end up with a kind of uncontrollable and chaotic environment. 

Now, at Dell, we focused on very heavily on internal governance, as you know. We focused on the top-priority projects. We had a business-first approach. We really were disciplined about picking the right things to go after and it’s paying significant dividends. But, generally speaking, areas we focused on were ones that didn’t have a lot of regulatory and governance frameworks on them from the outside so we could move fast. The minute you step into an area that’s heavily regulated or has a lot of third-party interest in it, it becomes really hard to navigate governance. There are techniques to do it. We are using the NIST framework as one way to approach it that, regardless of the government framework, if the NIST and EU frameworks are kind of the basis of it, you’re probably OK, but that’s not a sufficient answer.

So my first prediction is: Something has to happen to make governance better, and it’s going to become a significant priority both inside of the enterprise and at the governmental and national and international level. But the corollary to this is the call to action that governments and enterprises need to start working much better together to focus not on creating arbitrary governance frameworks, but asking, “Are we creating the right governance frameworks that allow us to move at speed as this technology evolves and get it to value quickly?”

Getting this wrong means that we create an impediment about improving the productivity of the world — curing diseases, solving massive problems. Getting it right — if we put the right governance and the right controls, but they’re orderly and they’re implementable and they take into account that actual companies have to go do this stuff — will allow us to create a sustainable, long-term framework.
So, item number one: Governance, governance, governance is going to play a very big role this year. It’s a mess right now. There are ways to make it better, but it requires everybody — from the governmental entities to the enterprises that want to use it, to the providers of AI — to start working much better together. If we get governance right, things are much better than if we have a chaotic, fragmented and incoherent governance framework. 

Prediction #2: Data management — the true backbone of AI innovation 

Number two is really about data management. We have learned now, as we’ve implemented AI, that the AI compute part is very important. And we know that which model you use and what GPU you run on it, you cannot do AI without all of that stuff. You know, having a place to compute AI is foundational. But it turns out that, in the enterprise, we’re not doing arbitrary AI, we’re doing AI that is actually tied to the unique information and data that exists within the enterprise. And in order to make that work in an enterprise context, you need to actually take that foundational information that lives in your systems of record and make it consumable by the AI world. 

Now, what that physically and technically means is: You transform that raw information into mathematical representations, or logical or graph-based representations that are suitable to power things like agents and chatbots. We are now describing that layer of technology that sits between the systems of record and the AI compute environment as “the knowledge layer.” That’s an industry term that’s emerging. What it says is things like vector databases, graph databases, knowledge graphs — these are all things that exist within the AI world that didn’t exist outside of it, but they are necessary.

More importantly, that “knowledge layer” is not a cold storage layer — there is no concept of cold. It tends to be highly transactional, it tends to be high intensity, it tends to be always hot or warm. And the result of that is that it has a strong affinity to where you do AI compute. Now, that’s fascinating because when you start thinking about where you put the knowledge layer, if you put it where your systems of record are, but your systems of record aren’t actually where the AI interaction is happening, you may be putting it in the wrong place. 

On the other hand, if you put your knowledge layer close to where the compute occurs, out in the real world, in your factories, in areas that you can control, you may in fact dramatically improve the performance, scalability, and security of your AI system.

Now, other consequences of the data management environment, it actually is probably the area that most valuable information of your enterprise exists. Take, for example, agentic. If you have knowledge graphs that are being used to create long-term memory and digital skill representation as your agents do work, they are actually producing this insight by processing all of the information that exists in our systems of record, and turning them into knowledge graphs and paths through the knowledge graph that actually are the ways to do work. The digital skills, the new processes that the AI systems figure out, that is incredibly valuable. If I told you I could run a set of agents on a set of data, and the agents would ultimately figure out the absolute best way to do 100 skills that are important to your business that you’ve never thought about, and where that was represented was not arbitrary; it was in a knowledge graph which is a set of data You would probably say that data is pretty valuable, and so we're gonna have to learn how to make sure that it's survivable, make sure that it's-- you can back it up, you can recover if there's a cyber event, nobody can steal it from you.

And so this whole layer of knowledge management and kn- the knowledge layer of the AI world, which is really the data management necessary to do AI, is an emerging space that in, in twenty twenty-six we expect to be a far bigger part of the dialogue. I think we've kind of figured out how to do AI compute.

I think we know how our systems of record work. What we don't really understand and haven't really matured on is how do we deal with the bespoke knowledge that is both used by AIs, but also created by AIs to be able to engage with an enterprise. And that's gonna be the high-value piece. That's the area that probably will attract the most attention in terms of new activity as we go forward, especially as we move into agentic.

Prediction #3: Agentic AI — the new operations continuity manager

Number three is, is agentic. I mean, last year I predicted that agentic would be the word of the year in twenty twenty-five. I think I got that one right. Uh, you know, and it turns out everything this year has been about agentic. But I do think there's, uh, you know, now a lot of learning about what are agents, and more importantly, what are they gonna do?

And today when we say agents, most people think fancy prompts, a little bit of automation, but that's really understating what they're capable of. Um, you know, it-- when we think about an agent, by the way, we're not thinking about a fancy prompt or a chatbot. We're talking about a new technology, and generally, we describe it as a software system that's capable of independent action and interaction with the real world.

In fact, when we take it apart technically within an agent, however you implement it, there are really kind of four capabilities that exist inside of it. One is, they are ba-based and powered by large and small language models, and what those models give them is general purpose knowledge, communication skills, and in many cases, reasoning capability.

That's a very important component, but an agent is not an LLM. An LLM is a component of an agent. The second component are-- we describe as, uh, today more knowledge graphs, but really what they are is new data structures that allow us to have long- and short-term memory to take custom and proprietary data and share it with agents so that the combination of the general purpose knowledge in the large language model and the very specific stateful information and knowledge that's tied to a specific skill in the knowledge graph actually work together to kind of power the brain of, of the agent.

But in addition to that, there is a-- there are two other components. Agents have one capability that chatbots typically don't. They can interact with the world outside of them. They can perceive by interacting with data sets, and they can act by actually using tools. And sometimes those tools are an IT tool, but they could be controlling a robot or any number of things.

Today, there is a protocol that's emerged called Model Context Protocol or MCP that is kind of the default way to do that. But we believe that a properly implemented agent not only has a bunch of knowledge and knowledge graphs and large language models, but it also has interfaces with the real world through MCP that allow it to interact with external data sources and control external tools, which makes it quite powerful compared to a traditional chatbot that just emits information as text in a window for you.

And then the last component is agents have always had an advantage over singular AI tools because architecturally they can work as a team. But the way they work as a team is because they have protocols and communication interfaces. Today, the emerging protocol that's kind of dominant in that space is something called agent-to-agent or A2A.

And when you put A2A into an agentic environment, the effect is now agents have a secure, understandable, standardized way to actually share information with each other to communicate, not just within an ensemble, but between organizations and between different agentic frameworks. You take all of that together and you have a very, very powerful tool.

But the prediction isn't that. That, that's, that's, that's just what it is. The prediction is when you understand that these are no longer just a tool, they are an entity that can actually do work, work as a team, interact with the real world, be very, very specific about what they're doing and keep state over the long term with memory, what is their role in the enterprise, and is it going to evolve?

Today, most agentic systems are used kind of as single tools. They do a task in isolation, and they're-- they maybe are a bit more autonomous. Fast-forward into twenty twenty-six, and what we see is that these agents have not only the ability to do specific work, but they have the ability, because they can interact with the world around them, to act as coordinating functions to help us collectively work better.

You know, and one of the things that we've seen that's fascinating, that is a prediction of this year, is that up until now, we viewed agents as a tool that humans would use to make the human work better And what we're finding is that agents, if done, if implemented properly, because they have an ability to keep state and, and, and operate over a long term, if you use them not just to do work for humans, but to help humans work better together to kind of coordinate their activity, to make sure that that handoffs or shift changes occur, are done properly in a factory, or to make sure that, that a long, complex process is understandable by an entire team, and that if something goes off the rails, the agent can assist the team in doing their job better, we start to see this kind of inverted relationship where a-agents may in fact not just be a tool for people, but they may be a capability that allows people to do their human jobs better.

Now, this is something very new, but what it exposes is we shouldn't think of agents as just a tool underneath humanity. We should think of them more as a partner and a, and a participant, that when we think about the collective work of people and machines working together, don't assume that there's a rigid hierarchy.

The agent may play a coordinating role, it may play a subordinate role and be a tool. The people may be able to describe intent and validation, which they will always do, but in terms of doing the work, they may actually define the work and then have the agents do it for them. Or they may, because some of this work has to be done by people, get input from the agents about knowing when to do the work and how to do the work.

And so it's a very complex relationship forming, and we expect it to be a pretty interesting topic this year as we think about how agents work with humanity. And what we're seeing already is that it is not as cut and dry as they are a tool for people. They are a partner and they are a participant in one of the components in a composite of work, which is the combination of people and machines working together to get to better outcomes. 

Prediction #4: AI factories redefine resiliency and disaster recovery

The, uh, the, the fourth prediction is kind of an obvious one, but we haven't done it yet, which is surprising, and I think there's going to be a significant shift in thinking about AI factories and where you do AI from just getting them running to making sure they always run.

Today, when we think about, you know, disaster recovery for an AI factory, it isn't even really a topic in many cases, or it's doing it in a legacy way. But an AI factory is not the same as a ERP instance or a CRM environment. It requi- it's a very different system. It's dependent on GPUs, it has a knowledge layer, it does different things, it has different degrees of value.

And so one of the things that's not been clear to date is what is the way to make an AI factory environment survivable over the long term, make sure that it never goes away? One of the most fascinating things about it, though, is when you think about resiliency of the AI factory, if you go back to a couple of slides ago, uh, you know, it's not that dependent on the systems of record.

If you've taken all of your support information and vectorized it into a knowledge layer, and then the AI compute layer is actually using that knowledge layer to provide services to your customers, there is no interaction for the most part or a very limited transactional interaction with the system of record for support to happen now.

It's more of a passive participant, the systems of record, and the active participant is the AI compute and the knowledge layer. Well, that has a consequence when we think about resiliency, because if I wanna survive a short-term outage, all I really need to do is think about making sure my compute exists somewhere and my knowledge layer is available and whatever minimal transactional interfaces occur.

But I don't have to have a backup of my entire enterprise to keep my AI systems running. That will open up tremendous opportunities to think differently about cyber recovery, about data resiliency. Uh, and my-- the takeaway from this is do not assume your traditional approach to data recovery and data survivability and disaster recovery and cyber resiliency are exactly appropriate for an AI factory and an AI environment.

They are a different architecture, they have a different set of capabilities, and they have a different set of demands. So I think that as we obviously are putting more and more AI into production in enterprises, we have risk associated with them failing, and the result will be that we'll spend a great deal of time this year probably innovating quite heavily on applying new concepts of data recovery, disaster recovery, resiliency to the AI factory world, and that's likely gonna materialize as entirely new architectures or new business models that come into play to make the AI world resilient in all scenarios.

And that, that I think will be very exciting. 

Prediction #5: Sovereign AI accelerates national enterprise infrastructure

And then number five is sovereign AI, a term that didn't exist, let's say, a year ago, but became very popular in the second half of 2025 about governments developing a strategy to be part of the AI discussion. Now, we were very vocal about this to say that, look, it-- there's many ways for sovereign AI to manifest, at least three major ones that you could have, you know, AI being built for the purposes of running a government.

You know, AI, you know, AI for government, uh, or government for government in this case. There are other approaches where it could be government for industry, where a government has a strategic strategy to build AI infrastructure but not use it for government, make it available to their industrial base to make sure their industrial base does well.

Many companies don't have the resources to build the at-scale infrastructure they need. If a government builds it for a country and lets their industry ac-access it, you can have breakthroughs in scientific and engineering and per-- and even manufacturing or petrochemical performance in a country. And then there's a third one that we knew a long...

that we know also was happening, which is government with industry, which is really this idea of the government helping industry organize and make sure that the industries in that country are advantaged as they go on the AI journey. Now, that's all old news. But as you think about this year, that, that theory of how will sovereign materialize is now becoming very real I, I predicted last year that, you know, by sometime this n- this 2026 timeframe, almost every enterpri- or every regional government in the world will have a sovereign AI strategy.

That is very much happening right now. Um, but once they have it, what occurs? What's the impact? Is it just better government services, or does a lot more happen? And so our prediction is that as sovereign AI strategies come into play, as people actually have their first generation of building some infrastructure, developing a policy, now they start to figure out what to do with it.

And we've thought about the different things that could happen if you have national infrastructure to enable the AI within your country. There are things like interworking zones, like if you have a billion agents in a country and you wanna make sure that they work safely with each other, there's going to be places where they, they coordinate, they connect.

Uh, that n- if you look at telecom exchanges and things of that nature, they tended to have national demarcations. There's no reason to think that won't happen within sovereign AI infrastructure. Fine-tuning as a service, this idea that, you know, we, we, we talk about large scale training, but most enterprises are now realizing, especially on AI PCs and in very specialized use cases, fine-tuning is important.

But fine-tuning requires a lot of infrastructure, and it requires it to be available for a short period of time in an organized way that's very trustworthy. And we think that sovereign infrastructure may turn out to be a very useful way to at least enable more of your industry, maybe even your small businesses, to create and maintain fine-tuned models over time if they're tied to something that has a national interest.

AI factory replicas, you know, disaster recovery. Why not have an option to do disaster recovery of critical infrastructure in an en- in a, in a, in a country be tied to infrastructure and AI strategy that's part of the sovereign AI strategy? Could be data centers, could be just governance, but it's very likely that government will play a role in the disaster recovery and survivability of at least their critical infrastructure industries, but maybe broader than that Um, robotics.

We see robotics happening at scale now, starting to accelerate dramatically, sometimes called physical AI. That's fantastic. It's very important, but, you know, robots don't operate in isolation. In fact, the best robotic environments in the world, we've seen this with autonomous driving, have world brains.

They have a place where you put a aggregated view of the, the world, not just what the robot knows locally, but what it, what could it know generally. In an autonomous driving situation, it would be the map of every road in the world. You don't need that on every car. And when I say map, it's a detailed understanding down to the pothole so that the adaptive suspension can work.

You couldn't possibly put that on a single car, but having it in a place that's accessible is very important to robots. Now, autonomous driving, that's being done by the au-automobile companies and relatively successful. But when you start to apply robotics to other areas like, uh, you know, someti- uh, the, the manufacturing sector, the even military applications, healthcare applications, the idea that the world brain ought to be somewhere that is tied to the jurisdiction it operates in starts to make more sense.

And so not all world brains will sit on sovereign infrastructure, but some of them, again, tied to the national interest, might need to sit on an infrastructure or environment or be part of a sovereign strategy so that the country that's using them feels comfortable that they can trust them. And then finally, you know, agents themselves.

Agents are not going to be kind of ephemeral, one-off, uh, you know, random technology. They're likely to have very long lifespans. So you may have agents that their knowledge graph accumulates information for years or maybe even decades as it gets good at skills. And the reality is, where you put that, where it lives out its life may in fact involve sovereign infrastructure.

It won't always be sovereign infrastructure, but I sometimes use the example of, you know, if you have a, an agent that's good at accounting law in a particular jurisdiction, the thing that makes it authorized to be an accountant in that jurisdiction is because the government has certified that they, that, that human being in the past and now that agent has the right skills, follows the right rules Well, that's all good, but if that agent lives in a different country physically, that may be a problem.

And so we do see that sovereignty of where the agent lives, not necessarily where it runs, but where its data is stored, where the knowledge graph lives, how it's governed, are going to be part of this dialogue. Now, don't take any of this as all things are gonna live on sovereign infrastructure. But when you look at a country and its overall AI strategy and what matters to any jurisdiction, there are certain things like critical infrastructure, defense, some research sectors, government operations that actually do have much more sensitivity, and there is a long-term bias towards wanting that to run in environments where the government has some control over.

We expect that whether it's agents, robots, disaster recovery, or even just running AI systems, training them and operating them over time, portions of those industries are all going to emerge as new use cases that could sit on sovereign environments, which will create an acceleration of not just sovereign AI, but the enterprise use cases and dependencies on them.

So these two areas are not separate markets. They're going to become more commingled, especially as, hopefully, we get our regulatory or, or, or regulatory regimes in order. Uh, and then the last thing to leave you with is we talk a lot about AI, but the bonus prize is keep your eye on quantum. There have been significant quantum breakthroughs in 2025.

We are seeing, you know, 100 qubit systems that are highly error-corrected, that are actually doing interesting things. We're seeing software systems that can reduce dramatically the gate count necessary to do advanced algorithms. We aren't at quantum supremacy or really even at quantum utility, even though there's some early examples of that, but we're moving fast.

And if you look at the curve in terms of qubit count and in terms of the technology, we've always said that quantum was not an if, but a when. I'm not sure the when is 20 Thanks. But we're gonna see continued improvement and some early indications of use cases that we can either use to sanity check existing approaches, which we've done in Dell, or to maybe solve very early problems.

But keep your eye on it, because I stand with my prediction I made a few years ago. When quantum becomes viable, which is not an if, it's a when, it will be a bigger disruption than what went on when, in November when ChatGPT came out. Because what will happen is the ability to do certain types of math that are very central to things like AI and, and the way we experience the physical world will suddenly become orders of magnitude more effective and capable and cost-effective, and that will transform whatever the state-of-the-art is that day.

So I don't know when that day is coming, but when Q day comes, it's gonna be a big deal. And 2026 looks like it's gonna be a pretty active year in quantum, and we may in fact see some early indications that give us more clarity about when that's gonna happen, but more importantly, what's gonna happen when it occurs, which is what we all have to start thinking about now.

So I'll, I'll end there. Um, you know, these are very specific predictions. They're not, you know, AI is gonna be good and agents are gonna happen. That's all old news. But I do think that what we're seeing is a maturing of the AI world. We're seeing much more specific technologies, much more specific problems materialize, and all of that will probably be a big part of the dialogue as we go into this year because people are trying to put this into production.

And when you make something go into production, you care about the details. You care about how to make it resilient, how to make it scalable, how to deal with the government environment, how to cra- navigate regulatory environments. And that's actually a very positive sign because fundamentally, you know, uh, two years ago, we were all talking in platitudes and high level.

Last year, we started to see early examples of early implementations, and now we're seeing a maturing, and that means that interest in these more detailed topics is likely going to accelerate. And that will probably characterize a lot of the dialogue in 2026. So, um, you know, we'll, we'll see what happens.

Uh, but I, I do think we're-- You know, one thing I was absolutely certain, 2026 is gonna be even more active and exciting than 2025 was in the AI world. And, you know, if you're, if you like that kind of thing, which most of us do, uh, it'll be a fun year. It'll be a busy year, and hopefully it'll be a very productive year for everybody.

So again, thanks very much and, uh, you know, let's see what happens with 2026.