Dell exec: Balance safety and speed with agentic AI ethics

What enterprise AI agents actually are

Roese began his talk emphasizing one of the simplest truths about agentic AI: AI agents aren't chatbots. This is a common misunderstanding that, in Roese's view, can undermine an organization's strategic planning.

"[Agentic AI] is a technology that is focused on the digitization of work," he said in the conference session "Agentic AI: Unlocking the Power of Purposeful Intelligence." "It is autonomous, it can reason, it can use tools, it can have memory, it can have specialized knowledge, it can interact with other agents. Do any of those things sound like a chatbot?"

Businesses can only begin to understand the human and organizational dynamics of agentic AI, he continued, if they first understand the difference between productivity tools with autonomous entities capable of reason that can do work. Agentic AI can shift many organizational norms: it can optimize workflows, shift human and agent roles within the organization, and require new security and ethical considerations.

Further, businesses need to understand the nuances between agent types to achieve optimal results and measurable ROI. Roese shared the following framework for categorizing the operational functions of AI agents:

• Low autonomy, simple work: Simple productivity tools.
• High autonomy, simple work: Can complete simple tasks with almost no human involvement.
• Hygiene agents: Autonomous entities that can fulfill objectives and use many tools.
• Coordination agents: Complete complex tasks without much autonomy (for example, monitor complex workflows).

 "It gave us some guardrails to correlate this technology to the thing that it was doing," he said.

These guardrails enable businesses to see results faster by requiring them to be deliberate about the agents they deploy, their scope and the intentions they have for them. Deploying the right agent for a specific task ensures efficacy, saving time and money while producing results quickly.

Faster time-to-market and the knowledge to effectively command AI tools will be the key differentiators for businesses going forward, according to Roese. Now that anyone with any level of technical experience can copy or generate software code with an open-source AI tool, products won't set organizations apart.

"The only real sustainable source of differentiation from a product and technology perspective in the software world is the speed in which you can execute," he said. "The intellectual capability to know what to do is super important, and that's a huge asset, but if it's not coupled with a highly agile, fast-moving ability to take your idea into conception, you will largely lose."

But speed without guardrails could lead to disaster. Almost paradoxically, it's because of standardized ethical and governance frameworks that companies like Dell and IBM can continue to innovate with the agility they need to stay competitive.

A protocol-based approach to agentic AI ethics

Roese said that the conventional way enterprises maneuver policy about trust and risk needs to change. The original model often involved many security and legal experts crafting policies to protect their business. But Dell thinks they can remove some of the cooks from the kitchen with more standardized systems.

"If every AI project is a random adventure where you pick your own tools and do your own thing, there's no structure -- then you need a lot of lawyers and security people to make sure you don't do something dumb," he said. "But, if you deploy standardized platforms, have a standardized operating model, have a proper governance structure … then have trained people who know how to do this, you can relax some of those things."

Roese said that Dell is about to roll out a new principle centered on trust: If trained developers are working on trusted, approved platforms, they will be given "a lot more latitude to move a lot faster."

IBM fellow Kush Varshney agreed with the idea that standardization increases trust and mitigates risk.

"Development toward predictability, standardization and robustness in models and runtimes is development toward trust and risk mitigation," he said in an email. "These efforts work hand in hand and actually speed up getting quality products and services to market."

Roese shared some key guidance on developing such deeply trusted agentic AI systems, such as the following:

• Establish clear rules for agentic use and the agent's boundaries.
• Issue digital identities for every agent.
• Work in safe, predictable, controlled environments.

He said businesses need to start thinking about protocols for probabilistic autonomous systems rather than policy. By using strong protocols to ensure agents perform sufficient reasoning to guide their processes, and by using tools to supplement them that provide greater predictability, the focus can shift to a more standardized, granular level rather than on the overarching, higher-order policies.

Organizations still need many layers of governance, "both intrinsic governance within agents … and extrinsic governance that wraps agents in guardrail infrastructure," Varshney said. This governance, he said, needs to be both organizational and administrative.

Varshney said that another major ethical consideration businesses need to think about is values.

"One of the underlying assumptions in much of today's AI development is that we can build a single, highly capable model and align it with a broadly defined set of human values. But in reality, values are plural, context-dependent and evolving," he said.

He said that developers shape these values through choices about data, objectives and trade-offs. Like Roese, he believes that one of the most important facets of agentic AI ethics is having trusted, educated individuals who manage and work with agents.

Everett Bishop is an associate site editor for Informa TechTarget's AI & Emerging Tech group, covering AI, quantum computing and other emerging technologies.