S&P Global's AI Lead On The Three Pillars for AI Success: Agents, Integration, And Trust

*All opinions expressed in this piece are those of Vishal Agrawal, and do not necessarily reflect the views of his employer.

While many enterprises are celebrating early AI wins through simple and road-tested implementations of chatbots and productivity tools, a much deeper and more disruptive transformation is already underway. The real leap isn't in building simple tools, but in rearchitecting entire business systems around autonomous agents that can act. This shift is happening at a breathtaking pace. As one industry leader put it, when experts predict a three-year timeline for a major AI breakthrough, it will likely arrive in six months.

But the finish line isn't just about speed. It's about building a new foundation of control, integration, and trust. For leaders navigating this landscape, the challenge is to look beyond the immediate hype and architect for a future where intelligent systems are not just a feature, but the very fabric of how work gets done.

We spoke with S&P Global's Vishal Agrawal, who is focused on leveraging AI-based machine learning to transform and automate complex data processes in capital market applications and beyond. Drawing on a career that includes serving as Senior Executive Vice President at Broadridge—where he managed global technology initiatives supporting $70M in revenue—Agrawal offered a perspective on what it truly takes to succeed. He argued that enterprises must master three foundational pillars: moving from simple bots to governed agents, conquering the hidden complexities of system integration, and building unwavering trust through transparency.

To truly harness the power of this new wave of AI, leaders must first understand what separates a simple bot from a true autonomous agent. Agrawal offered a clear definition.

• The three qualities of an agent: "An agent has three important characteristics: one, it can perceive; two, it can reason; and three, it can act autonomously. Those three qualities are what differentiate an agent from a bot."

• Where risk resides: This capacity to "act autonomously" is where both the immense value and the significant risk reside. Agrawal explained that enterprises operate on two types of workflows: indeterministic ones like content summarization, where outcomes can vary, and deterministic ones, like fixing an IT issue, where actions must be precise and repeatable. It is in these deterministic, high-stakes environments that the danger becomes most acute. He pointed to a cautionary tale where a poorly governed agent deleted an entire company database—a visceral reminder of what happens when power outpaces control.

To prevent such cascading failures, practical frameworks are emerging to enforce control. This starts with putting guardrails in place at three levels: ensuring the agent's reliability, securing it from cybersecurity threats like prompt injection, and embedding ethical oversight to eliminate bias.

• Checks and balances: "There are models based on the 'maker-checker' principle. There are agents which make the call, and there is another agent which validates that call. If they don't agree, it goes to a human to validate. But I have seen very little evidence where agents are making critical decisions. They are semi-autonomous. If there is a critical decision, you ask the agent to get a human involved to make those decisions."

While the agent itself is the star of the show, Agrawal argued that the hardest part of the enterprise AI journey isn't the model; it's the integration. To deliver real value, agents must connect to a complex web of legacy systems, APIs, and internal knowledge bases like Jira, SAP, and Salesforce. This is the next frontier where the most significant gains will be made.

• The integration challenge: "The 'could' part of it goes back to ROI in most cases. What do agents need? They need clean data. If the data is not clean, we know that agents can make mistakes; they can hallucinate. To truly unlock the capability of the goldmine of data you have within the organization, you need to be able to access the systems."