Breaking down the complexities of Agentic AI

“You need to have a very clear accountability chain. If something is not right, who takes the accountability? You can't go and blame the agent and say the agent misunderstood what you're saying,” says Soo Mei May, Chief of AI Global Solution Specialist, Dell Technologies.

As businesses around the world look towards embracing Agentic AI in their organizations, there are still those that are unsure how much of an impact the technology can have on their business despite its hype. Apart from discussions of ROI, understanding how Agentic AI works, especially on the infrastructure needed to enable it to remain key talking points.

For Dell Technologies, Agentic AI will be the biggest paradigm shift to happen this year as enterprises look beyond automation. The maturation of GenAI given rise to sophisticated AI agents capable of autonomous operation, natural language communication, and seamless collaboration with both humans and AI agents.

At Dell Technologies World 2025, Dell is expected to showcase more use cases on how Agentic AI is making a difference to organizations from various industries around the world.

[Related: Dell Technologies expects a more focused AI era for 2025]

In an interview with CRN Asia, Soo Mei May, Chief of AI Global Solution Specialist at Dell Technologies breaks down what the technology can actually do for businesses as well as the infrastructure that will be required to deliver the best results.

What's the best way to define what agentic AI is?

Agentic AI from Dell's point of view is a software system that uses artificial intelligence to make decisions and take actions to achieve a set of objectives. It is a combination of many things.

One, it perceives its environment, which is it's able to collect information from the environment. Two, it processes that information, and it is able to make decisions. This is where the reasoning and the decision making happens. And three, it can take actions to achieve a certain goal. Four is an option which can be learning or adaptation. You can program that in, or you can just leave that out. It's optional. So, the main components are there.

Now, if you take an AI system, it is a combination of many things. One is there's a large language model involved. Two, there must be a system prompt. System prompt is where you define what the AI agent is and what it can do, what it cannot do.

The third one is the controller module or the planning module, where it does all the decisions on executing tasks. The fourth one is memory. Memory is very important in the agentic AI system because it has to remember everything that we talked about and then be able to store it in a safe place and then retrieve the right things for continuation of conversation.

The fifth part is the loop or the orchestrator. You've got to have that loop. This is where the agent is thinking, evaluating, and taking action until it decides that, okay, it is good enough and the objective is achieved.

And then there is the tool interface. So, the tool interface is where you have to interact with an external environment, like a web search, file search or API tools. The last one would be the interface with the users, as well as connectors with the other agents. That itself is a whole agentic AI system.

What are the investments that are most critical for businesses today when it comes to having agentic AI solutions?

For an Agentic AI system to be holistic or fully autonomous, businesses need to think about investments that are end-to-end, which is broader and not just pure data center or pure public cloud or edge, but across a whole range.

For example, chatbots can become an AI agent in a way. But today, chatbots are all sitting in the data centers. Now, what if I have this information in my laptop? And perhaps I want my orchestrator to be in my laptop instead of in the data center so that I can distribute the workloads? And how would that happen if we don't think about that whole end-to-end ecosystem itself?

That's why when it comes to infrastructure investment, it is not just to invest in this or invest in that. It must have that whole flow of what your Agentic AI is going to do for you and where it's going to execute.

Is it going to be just pure data center? I don't think so. It will be everywhere. So investment will be across the whole portfolio of data center, the cloud, your private cloud, the edge, as well as your local machines.

With all these investments, how can organizations measure their ROI for Agentic AI?

Measuring ROI is extremely difficult. That is why we advocate that all customers focus on the functions that matter. For Dell, we focus on the four functions that will be directly impacting our revenue and our top line. The four functions would be the sales, the services, our product engineering, and of course in our supply chain.

Invest in these because these are the ones that are going to be impacting your revenue directly. And then if you want to go and measure your ROI, of course you have to do your use case analysis, which is feasibility versus your value. Then you focus down to the ones that bring in the highest score in terms of feasibility and value.

In Dell, we've got like over 800 use cases, but we didn't go after all of them. In the end, we narrowed it down to about 100 and we broke it down into six archetypes. And these six archetypes are not all GenAI. It’s a split between GenAI, traditional AI predictive analysis and business intelligence. We know that these are the ones that will impact and will give the best ROI because we've been measuring them from the feasibility and value point of view.

So, when you know what to focus on, will that also impact how you invest? Is the infrastructure for GenAI different to just AI?

There is traditional AI, which is all predictive analytics. That one is very different from GenAI. Now GenAI, you expand it further to Agentic AI is actually additional expansion. It's not too different, but it will be upgraded.

For example, in a very straightforward chatbot, like ChatGPT, where you ask a question, that's an input token. And then it answers, that's an output token. And that's it.

But in an agentic AI system, you've asked a question, it’s an input token. Then it goes to an orchestrator. Orchestrator will need to pull a tool to get some answers. So, this one becomes an input token to another section. And then the output token is here, comes back, and then it goes to the next one. And the agent needs to refine this further.

Again, it goes to another agent. So, another input to another agent and it becomes continuous. So, at every step of this system, there is input-output, input-output, input-output.

What does that mean? That means that it's very compute-heavy. So, if previously your chatbot requires just, let's say, two GPUs with 80 gigs, today with Agentic AI, you probably need more than that. One of our distinguished engineers says it could result in 20 to 30x of tokens increase.

In the recent NVIDIA GTC, they actually said an increase of 50x tokens. So just from that perspective, your compute may have to go up. So, you're talking about more servers. Then you've got the persistent memory that you need. We're talking about more storage, as well as security in your storage. So, these components will all have to come in. I won't say it's different. It's just that it's going to be upgraded extensively.

What about the ethical considerations when developing Agentic AI, from privacy to guardrails?

You have to be even more careful than GenAI today. Because remember, your agents are autonomous, meaning you're giving them the power to go everywhere and do anything without human intervention.

If you do not have that visibility, it is not good. Transparency and explainability are very important. If you think about the development of the model context protocol, it is very exciting because you're not going to need orchestrators. It's going to make decisions very autonomously within the model itself. But then you wouldn't be able to track it. The logs will not be clear.

And that's why when you have Agentic AI systems, you need to have proper logs. For example, API calls, why are you calling them? And you need to keep the calls under control as well.

You've also got to make sure that there is consent from the users, that they know that you are going to be storing the chats in the memory, and they consent to that, they agree to that. So that privacy side needs to be there. Plus, they need to have the right to delete, so that you don't store this in their memory.

You also need to think about, from the perspective of how it makes a decision. Is it fair? Is it biased? In the chatbot itself, it's already very hard. So, what more in agentic AI systems, right? Let's say, it will fill up an application form for you, when you're applying for something, and then maybe it will make a decision for you as well. The explainability has to be very clear.

And the most important thing is accountability. You need to have a very clear accountability chain. Your agent is so autonomous, it's going to decide to do this and that. If something is not right, who takes the accountability? You can't go and blame the agent and say the agent misunderstood what you're saying. No. Someone has to take accountability when something goes wrong.

So that accountability chain is a human, actually. You need to know what went wrong, and who needs to be accountable for it. So that's why, with agentic systems, it's even more important that you pay attention to these guardrails, compared to your large-language model application, because of the nature of its autonomous decision-making.

Would that make businesses think twice about implementing it, because of this nature of how it is?

I don't think so. I think they are very excited, because they can see that when things can run autonomously, businesses are thinking, oh, it can really help save a lot of time, a lot of cost. That's why they're going to go for it.

Also, because they're aware of this, they're going to be putting guardrails and pay attention to what needs to be done and executed before they even run full steam ahead.

Given the high compute power required for Agentic AI, will the need for infrastructure only get heavier and where does sustainability fit in?

I do think it will get heavier. That's why Dell is working very hard in building a data center, considering the cooling system, the impact and design it in a way that uses very efficient cooling.

For example, our servers come with liquid cooling at the back. Of course, not all servers can go into immersion. But if your data center can support immersion cooling, that's the best.

Because it brings your PUE down quite low. Our PowerEdge servers have liquid cooling. And that's why when we go into GenAI and AI, we're already thinking about the energy capacity and the cooling. That's why it's built into these huge servers that's going to take a lot of energy and a lot of heat. And that's what we will do for sustainability.

Also, don't forget AI PC. So, it's not just the data center. It's also on your laptop. AI PC can handle some of the workloads, taking away the workload from your data center. So, if you can take it away from the data center, then just make use of your AI PC. It also helps in your conservation of energy as well as the cooling requirements.

Think about an AI PC as having a GPU or an NPU that can run your GenAI capabilities and have an orchestrator. So, the orchestrator doesn't have to sit in the data center. This means you don't have to keep hitting data center and APIs and then the compute is running crazily over there.

You can run the compute here on your AI PC. And if it's not enough, then you go to the data center. I feel that this will also help with taking away the workload from the data center and really help with sustainability.