Exploring the Future of Agentic AI

This blog is authored by David Nicholson, Chief Technology Advisor of The Futurum Group.

tl;dr: Agentic AI is advancing how people and AI collaborate by taking on tasks like gathering data, triggering actions and improving workflows. In this Q&A, Dell's John Roese and MIT's Pattie Maes discuss its real-world applications and challenges, as well as the future of human-AI partnerships.

Agentic AI represents the next frontier of innovation, offering powerful new ways to drive human progress.

To explore what this means for businesses and society, I sat down with two incredible minds: John Roese, Global Chief Technology Officer and Chief AI Officer at Dell Technologies, and Dr. Pattie Maes, Professor of Media Arts and Sciences at the MIT Media Lab. As host of the conversation, I had the privilege of guiding our discussion on the transformative potential of agentic AI, its practical applications across industries, and the future of human-AI collaboration.

The discussion provided valuable insights into how this technology can help people and organizations achieve more, and I'm excited to share those perspectives with you.

The following has been edited for length and readability.

John, what is agentic AI? And more importantly, what isn't it?

Roese: The world combines digitized information and skills. Agentic AI digitizes those skills, transferring work, reasoning, and behaviors into an AI system. An agent is not just a tool but something autonomous. You can delegate tasks or missions, and it figures out how to accomplish them. It interacts with tools, gathers data, and triggers actions in the real world. Unlike traditional systems, agents are proactive, observing and responding to their environment. This shifts humans from being "in the loop" for every decision to "on the loop," overseeing and guiding the agent's work. While traditional tools are valuable, agents represent a transformative new capability.

Pattie, John mentioned the idea of the types of interactions and collaboration that happened between humans and agents. As an expert in human-agent collaboration, what does this interaction look like, and where are we now in this collaborative effort between humans and agents?

Maes: For over 30 years, I've advocated for giving devices more autonomy to handle tasks on our behalf. As work and life grow busier, agents help manage information overload and multitasking. In both personal and professional contexts, people will increasingly rely on a collection of agents to assist with various tasks. However, early on, these agents won't fully automate tasks, as AI still makes mistakes. LLMs, which power many agents, can hallucinate or misinterpret the world. That's why it's crucial for humans to stay involved, audit agent behavior, and ensure they operate as intended, especially during initial adoption.

John, as we consider the journey toward autonomous behavior, it's clear that collaboration between humans and agents will play a pivotal role. Could you share which sectors or industries might face unique challenges in adopting this technology, and which are best positioned to embrace it? Pattie, I'd also love to hear your perspective on this.

Roese: I compare it to using training wheels. Agents will start with strict constraints, only handling tasks we're comfortable delegating. Over time, as we trust their behavior and data, these constraints can be relaxed, allowing more autonomy. This gradual approach applies across industries, though heavily regulated sectors, like software development, will adopt autonomy more slowly. In contrast, industries like telecoms are advancing more aggressively. Every industry has a starting point where AI can be trusted to handle specific tasks, and once that trust is established, expanding capabilities becomes much easier.

Pattie, do you have a different view based on the human equation?

Maes: Agents have a significant role, but we must proceed carefully. Starting small and gradually increasing their autonomy is key. People will always want to oversee their agents, track their actions, and audit them regularly. Maintaining trust and ensuring humans retain ultimate control over processes is essential.

What are some of the emerging technologies that we're all going to have to keep up with as humans that are going on right now?

Roese: The biggest challenge is making agents work together, which requires developing interworking standards. Current efforts focus on agent-to-agent communication, registries, and interoperability models. While today's key technologies are clear, this fast-moving space will continue to evolve rapidly, bringing new innovations.

Where are we going to be on this agentic journey by the end of this year, Pattie? What are your predictions?

Maes: I expect early business experiments with agents in low-risk tasks. Over time, agents will act as personalized assistants, managing tasks, automating processes, and answering simple questions, helping us stay organized and handle busy schedules more efficiently.

John, walk us through where you think we'll be by the end of the year and where we are currently. Help us understand where people should be starting today, and how to get started.

Roese: I've given this advice recently to colleagues and people across the industry. Three things in this order. First is to identify a task that's important but low risk if mistakes occur. Focus on areas that improve business operations without heavy regulation or high stakes. Second, define the agentic workflow. Map out what tools, data, and interactions the agent will need to perform the task. This is more about planning than technical work. And third, test the agent in a controlled environment using randomized data to evaluate its effectiveness. Don't overhaul existing workflows; instead, enhance them. For example, use agents to automate repetitive tasks like retrieving prices or inventory data during sales processes. Similarly, in software development, target areas with excessive human intervention. By the end of this year, I think the most mature agents in the real world will complement existing AI workflows, improving efficiency without replacing them. It's complementary to the first-generation AI workflows that we've already put into place.

From a "back to human technology" partnership perspective, do you see the starting place for individuals being different? I'm thinking of using agents to help myself become more efficient. Where would you recommend people start that journey?

Maes: The younger generation that I work with at MIT use today's chatbots at least two hours per day for solving all types of problems or questions. I think that generation will also be very quick with adopting agents that go beyond just providing information but also automating things themselves. I think they'll still be fairly constrained in terms of what kinds of things they are allowed to do.

Pattie, have any of your thoughts from 30 years ago surrounding artificial intelligence come true? What surprises have you seen?

Maes: After getting a PhD in AI, I shifted my focus towards what I call "IA" or intelligence augmentation. I didn't want to make computers and robots smarter; I want people to become smarter and more capable of learning. That's why I've been talking about software agents in 1994 and arguing that we should change the way we interact with our computers to help people understand that there is a technology implementing this, and it does have boundaries. The best solution is to have the technologist and the businesspeople at the table with them for the entire process to work better.

Are you more concerned today than maybe you would have been 10 years ago?

Maes: I am. For the last couple of years, the philosophy has been: if we can build it, we build it, and then, we throw it out there and encourage everybody to use it. And we're conducting an experiment in the real world as opposed to letting researchers conduct smaller scale experiments to understand possible problems. That's also why I recommend that businesses think carefully about what processes are not critical and are very repetitive in nature, providing a better opportunity to deploy carefully monitored agents and so on, ensuring that nothing goes wrong.

John, you talk to CIOs and CTOs consistently. Where within Dell are your professional services and solutions engineering folks really being pushed to the forefront?

Roese: Our PS organizations and our business transformation folks spend more time on the non-technical aspects of AI, helping customers figure out what process, where to apply this, how to safely apply this. And you still must do the technical work. The order of operations matters. And I will tell you it's inefficient to do only the consulting work because it must be built and implemented. It's also unwise to do the technical work without knowing what problem you're solving and how that problem should be best solved. This is just another example of once you do transformation, it's always an ecosystem, it's always multi-dimensional, it always has a human component to it. This is like every other technology currently. It's just faster, more impactful, and more interesting.

Do they need to understand any of the technology?

Roese: When we are evaluating or creating a path to do agentic, there is always a technologist in the room. This technology is not unbounded. It is not infinitely capable. It has risks. And if you're developing, even if you're picking the process, you should do it with a conscious understanding that there is a technology implementing this and it has boundaries. We should know you cannot hand off a purely intellectual concept into a technology ecosystem without ever connecting these two. Technologists and businesspeople sit together for the entire process.

This keeps going back to Pattie's area of expertise on the intersection between technology and people. This isn't an easy thing for only businesspeople and technologists to figure out. Would you agree?

Roese: It is incredibly likely that within the next two years, whatever you do today, whatever your work is defined as and the amount of effort you need to spend on it will change profoundly. And a huge portion of that work that could be automated could be delegated but can't be because there's no one to delegate it to. And you fundamentally scale. That's good.

You can watch the full conversation I had with John and Pattie here.