Five Questions Leaders Need to Ask about Agentic AI

AI is crossing a threshold from processing information to executing outcomes. On a recent episode of the Citi Institute podcast, the Institute’s Ronit Ghose spoke with Dr. Prag Sharma, Head of Citi’s Artificial Intelligence Center of Excellence, about agentic AI: what it is, how it differs from generative AI, and what leaders need to consider as these systems begin to interpret goals, make decisions, and even collaborate across workflows.

(The following is a transcript of this conversation. It has been edited for clarity and brevity.)

What is agentic AI, and how is it different from generative AI?

Ronit Ghose: We’ve all heard about generative AI over the last couple of years. How do you define agentic AI?

Dr. Prag Sharma: Artificial intelligence is difficult to define because it combines many different disciplines. Generative AI is one branch, where language models are designed to create new content.
Agentic AI is similar, but with an added layer: the ability to take action. It’s about combining content generation with the capacity to act on that output. That’s what makes it “agentic.”

Ronit: So generative AI shows us things—words, images, videos. Agentic AI acts on that output. Where does that apply in finance or the broader economy?

Prag: In many organizations, especially large financial institutions, consuming information is just one part of the work. Language models can summarize and extract information for humans to review. But acting on that information still requires human input.

Agentic AI goes a step further. It doesn’t just inform the user—it can execute the next step. That opens new types of use cases and opportunities for efficiency gains. Specifically, in addition to automating specific tasks, agentic AI is now enabling organizations to automate some end-to-end processes in their entirety.

How should organizations think about safety and control?

Ronit: If these systems are acting on our behalf, how do we manage safety and oversight?

Prag: It starts at design time. You need to build in guardrails, constraints, and interpretability from the beginning. You're not just triggering a task—you’re enabling systems to complete processes independently, which may not follow a predictable path.
These systems are often non-deterministic. The same input won’t always lead to the same output, so traditional QA methods don’t apply.

Ronit: You can’t test the same input and expect the same result.

Prag: Exactly. Instead, we simulate. We anticipate failure modes. We define escalation paths. Resilience comes from planning for what might go wrong—not just confirming what works. 
 

What role does intent play in how these systems function?

Ronit: You often mention prompt engineering. Can you explain what it means?

Prag: These systems rely on how we define objectives. If the task is vague—if we don’t specify what success or failure looks like—the system might optimize in unintended ways.

If you simply say “Do X” it might focus only on that metric, ignoring other important factors. That’s why we need to be deliberate in defining both goals and boundaries.
 

How will agentic AI change how we work with machines?

Ronit: We’re not just talking about automation anymore. These systems seem to collaborate and reason. 

Prag: Yes, they can ask clarifying questions, negotiate trade-offs, and operate across workflows. They’re no longer just tools—they’re part of the process.

That changes how we work with them. It’s not only engineers who need to understand AI. We need AI literacy across the business. Everyone should know how to interact with these systems, validate their reasoning, and work alongside them.
 

How should leaders think about impact and value?

Ronit: What should leaders expect from agentic AI?

Prag: It’s not just about speed or cost. These systems can simulate scenarios, explore alternatives, and test ideas, at a scale humans can’t match.

Ronit: So the value is in decision-making, not just efficiency?

Prag: Exactly. Some benefits are indirect, like improving judgment or creativity. But we have to measure them intentionally. If we only track time or cost savings, we miss the broader impact.

We should ask: Are we making better decisions? Are we surfacing more possibilities? That’s the real value.

Ronit: This is clearly a space where we’re just getting started but the questions we ask now will define how we lead later.