Agentic AI and what we should understand: A Q&A with Phaedra Boinodiris

Need to plan a trip? Book a train? Plan some meals? Agentic AI can do that and more. The intuitive AI system is sophisticated and designed to autonomously decide. But it requires robust governance due to its decision-making autonomy

The necessity for comprehensive governance structures grows paramount as these agentic AI systems gain popularity across organizations. These systems’ ability to decide independently underscores the critical need for robust frameworks that ensure ethical conduct, transparency and accountability. Phaedra Boinodiris, IBM’s Consulting’s Global Trustworthy AI leader, discusses the ‘hype’ around agentic AI and the need for foundational AI governance in this Q&A.

Teaganne Finn: Please explain agentic AI. Is the excitement surrounding this technology warranted?  

Phaedra Boinodiris: There's a tremendous amount of interest in agentic AI. If you think about it, it's like having a number of different AI solutions or AI models strung together, where the output of one AI model will serve as the input for another AI model.

TF: Could you share some examples of agentic AI in action?

PB: Absolutely. Let’s start with autonomous vehicles—a classic example. These systems go beyond just image recognition; they incorporate natural language processing, decision-making and even planning capabilities. They’re not just seeing—they’re reasoning and acting.

A more relatable example might be a smart home cleaning robot. It maps your living space, avoids obstacles, learns optimal cleaning routes, and decides when to clean based on your schedule or preferences. That’s an agent acting with autonomy in a physical environment.

Or consider a smart personal AI agent. Say you’re planning a trip to Madrid. You tell it your hotel location, your interests—music, food, art—and your constraints, like a budget of EUR 50 and a walking limit of 10 miles. In response, the AI agent curates an itinerary, makes reservations, and optimizes it based on time, proximity and preferences. It’s not just answering your questions—it’s doing the work for you.

Now let me share some enterprise-grade use cases that IBM has worked on:

In financial services, one agentic AI solution classifies and triages customer complaints. Using natural language processing, it identifies keywords and sentiment—like “refund”, “injury”, or “safety”—to prioritize and route messages appropriately. The system can even draft personalized responses aligned to company tone and policy. 

By structuring unstructured messages into clear categories and actions, the AI reduces resolution times and ensures that urgent matters are escalated for human review. IBM has implemented similar solutions with major banks to enhance customer experience and regulatory compliance.

In life sciences, agentic AI is used to streamline regulatory documentation. One IBM solution automatically generates high-quality draft reports from clinical trial data, research publications and technical documentation. Trained on previous regulatory submissions, the AI produces content in structured formats compatible with agency guidelines.

Human reviewers remain in the loop, ensuring accuracy and traceability. It not only accelerates time-to-market for new drugs but also reduces manual effort and error—which is critical in a highly regulated industry.

For appliance manufacturers, IBM has delivered AI-powered self-service tools to manage complex, long-tail customer inquiries—like finding replacement parts, comparing products or troubleshooting issues. These queries are diverse and infrequent, often too niche for traditional chatbots. The agentic solution pulls information from disparate knowledge bases to provide precise answers, reducing the load on contact centers and improving customer satisfaction at scale.

TF: With new technology comes greater risks pertaining to AI ethics and governance. Can you explain some of those concerns? 

PB: Predictive or traditional AI comes with risks that need safe guardrails. There are plenty of stories in the news that detail how organizations with great intent deployed inaccurate or unfair traditional AI models. Generative AI amplified those risks and introduced new ones. Now with increased autonomy in AI agents, there are yet new and amplified risks. 

For organizations that may not have foundational AI governance in place, those organizations that are looking to leapfrog into agentic AI have an even bigger haul. An even bigger learning curve when it comes to figuring out how we ensure that these AI solutions are trustworthy. 

To truly harness AI’s potential, organizations must not only “do the right AI”—choosing the right AI model use cases to invest in that are aligned to the mission strategy—but also “do AI right”. In order to “do AI right”, organizations must move beyond a focus on infrastructure, models and operations to address the sociotechnical challenge of earning trust.

Trust in AI systems cannot be achieved solely through technical solutions—it requires a holistic approach encompassing people, processes and tools. Trust must be intentionally built through transparency, education and inclusivity.

TF: A criticism when it comes to responsible AI governance is the stifling of innovation. Could you describe this balance of ethics and creativity? 

PB: I think that when people are critical about the subject of AI ethics and governance, or critical about the subject of regulation when it comes to AI, they have expressed concerns about the stifling of innovation and creativity. But I would point to the 75% failure stat that many analysts' references. 

75% of investments in AI fail. They get stuck. They get stuck in proof of concept. They never bring the expected ROI that investors are seeking.  And one of the core reasons why they fail is because they haven't earned people's trust or people are OVERLY trusting of the AI.  

The primary goal should be to choose the right AI to solve the problem that will augment our intelligence, and then to do that AI right by doing the hard work required to earn people’s trust. There is no easy button here.

AI systems need to be secure against adversaries. They need to reflect intent appropriately. They need to have fair outcomes, be transparent and truly explainable, and maintain our data privacy. Agents need to be observable so we can track what is actually happening. And yes, given the insatiable appetite that AI has for energy and water, we need these systems to also be environmentally sustainable. 

You cannot have trust without having accountability. I have written extensively on the subject of accountability, and we have released many IBV papers on this subject. Gartner just this year published an excellent paper detailing to their clients how to hold vendors contractually accountable for AI outcomes.  

And as I mentioned earlier, none of this is easy. There is no easy button to develop AI in a way that is worthy of our trust, but the effort is critically important. And once you do the foundational effort, THEN you can go fast as explained well in this race car track analogy. Especially as new technologies like agentic AI continue to evolve, more governance will be necessary.