Organizations need to design AI agents incredibility thoughtfully

“What I can say is organizations in the future will still be a collection of people that have jobs and do stuff, but the way they interact and the way they do that stuff is going to be completely and fundamentally changed,” says John Roese, CTO, Dell Technologies.

As organizations look to deploy more AI agents, there are concerns on the security of these agents as well as how much of an impact AI agents can have on the workforce. Over the past few months, there have been reports of how some companies have struggled with managing their agentic agents, especially when the guardrails are not properly designed.

According to John Roese, CTO of Dell Technologies, there is still a lot of confusion about what an agent is among organizations.

“The term is wildly overused, it's confusing, and there's a lot of apprehension. You hear agents can do good things and bad things, but if you don't actually know what an agent is, then anything that potentially could be an agent is potentially risky. So, the first thing is to just clarify and get a very crisp picture of an automated feature in an AI tool is not an agent. Unless it operates autonomously, unless it has the ability to reason, unless it can execute without necessarily human curation, it's not an autonomous agent, and you probably have a very different security model for those types of tools,” Roese said.

As such, Roese believes that organizations should have a pattern of using agents to accelerate the speed in which decision makers can access the information needed to make an intelligent decision.

“We're just beginning. What I can say is organizations in the future will still be a collection of people that have jobs and do stuff, but the way they interact and the way they do that stuff is going to be completely and fundamentally changed, and it will likely be quarters of a magnitude faster and more efficient than it could ever have been just throwing conventional approaches at it,” Roese said during a Q and A session at Dell Technologies World 2026.

For John Scimone, President and CSO at Dell Technologies, looking at the change AI agents are bringing to organizations today, there will be agents that are created by malicious actors for malicious purposes.

“When organizations are building their agents, they're not building them to be inherently malicious. At the same time, the nature of the data science is that one of the things that makes agents so powerful and so opportunity-creating is that they are absolutely relentless at achieving their objectives, typically by design. And oftentimes, if not well-managed, not well-designed, they can relentlessly pursue good objectives and actually create unintended and, at times, harmful outcomes,” he said.

For Scimone, it is imperative to design agents incredibly thoughtfully. This includes managing the way the outcomes are pursued and having guardrails against unintended malicious behavior.

However, when it comes to agents being manipulated by bad actors, it may not necessarily be from outside an organization. Agents can be manipulated by malicious actors who could be an internal employee or an external entity.

“To be able to manipulate agents to behave in unintended ways, even though the agents are not designed inherently as malicious, is also another avenue that needs to be contemplated. But these are all examples of why it's really important, from day one, as you're standing up an agentic political program, to thoughtfully substantiate zero-trust architecture, governance policies, and be really thoughtful on what types of work tasks agents are stepping into. There are inherently less risky work tasks that agents can be handed,” he added.

Flexibility and risk management for agents

As the agents get deployed, Scimone believes that flexibility can be slowly extended based on what the agents are stepping into. However, this is all a risk management continuum in terms of how an organization steps into the space.

“In my experience, I see a lot of conversations and concern around what could go wrong. The thing I'm most concerned about, honestly, is that people aren't talking about what could go right and how quickly can we do that for organizations. There's side things and for people to not step into it confidently and assertively, in my view, is a misstep,” he said.

For Roese, if agents are deployed into an environment that does not have good enterprise hygiene and good security hygiene, the probability of something strange happening and the agents stepping outside of boundaries because there aren't any and being enforced is quite high.

“In a well-structured enterprise with proper data hygiene, proper cyber recovery, proper policy, where agents are bound them to do a specific piece of work and not to do whatever they want, you end up with a much lower potential risk for those kinds of anomalies,” Roese said

“Agents are not independent of your IT architecture. They are dependent on it. And your IT architecture and security architectures are there to introduce protection and control so that if they're absolutely missing, you're going to have some strange stuff happening. But if they're there, it won't prevent anything from happening that's bad, but it definitely gives you a different experience with dealing with these technologies,” he concluded.