The rapid rise of agentic AI systems is reshaping expectations around accountability, risk management and governance across India’s financial sector. As autonomous systems begin to take on more complex decision-making roles, the industry is being compelled to rethink how trust, oversight and human judgement are embedded into AI-driven workflows. These themes took centre stage at the recent AI@Work: Shaping the Future of Business with AI panel discussion in Mumbai, moderated by Nagaraj Nagabushanam, Vice President, Data and Analytics and Designated AI Officer at The Hindu.
The discussion brought together senior technology and risk leaders from across banking, insurance and enterprise technology, including Rohit Kilam, CTO, HDFC Life Insurance; Premraj Avasthi, Head – IT and CIO, GIC Housing Finance Ltd; Pushkal Tenjerla, Head IT Security, RBL Bank; and Rajesh Malhotra, senior leader, Data & AI, IBM. Together, they examined how agentic AI is altering not just operational models, but the very foundations of governance in regulated industries.
Reimagining Governance for Autonomous Agents
The conversation opened with a close examination of how agentic AI disrupts traditional oversight structures. Kilam framed this shift as a fundamental change in the tempo of governance itself, observing that “we are seeing a transition from slower governance to a faster governance.” As autonomous agents are designed to act, learn and adapt in near real time, conventional post-facto controls are no longer sufficient.
To address this, Kilam outlined three governance models that organisations are currently navigating: fully autonomous systems, human-in-the-loop workflows and human-on-the-loop audits. Each represents a different balance between machine autonomy and human supervision. However, he emphasised that regardless of the model adopted, the principle remains unchanged: governance mechanisms must operate at the same speed as the systems they are designed to supervise. Embedded controls, continuous monitoring and real-time intervention are becoming essential features rather than optional safeguards.
The Enduring Relevance of Human Judgement
While agentic AI promises efficiency and scalability, Avasthi underscored that human judgement remains indispensable, particularly in financial decision-making contexts that demand nuance and contextual awareness. He pointed out that “there has to be a moderation in place that the AI has to learn the aspects of human intent,” highlighting the limits of automation in domains shaped by diverse borrower profiles, regional policies and socio-economic variables.
In lending and housing finance, decisions are often influenced by unstructured data, behavioural signals and situational factors that are difficult to codify fully. Avasthi argued that these complexities require deliberate human interpretation, even as AI systems assist in analysis and pattern recognition. Rather than viewing governance as a binary choice between humans and machines, he positioned it as a shared responsibility, where oversight is distributed across technology, process and people.
Trust, Risk and the Compliance Mindset
From a risk and security standpoint, Tenjerla broadened the discussion by linking governance directly to trust. He cautioned that the success of agentic AI in BFSI environments depends not only on technical robustness, but also on behavioural reliability. “We are not just governing the technology part of agentic AI here. We need to govern the behavioural part of it,” he said, pointing to the importance of defining acceptable system conduct alongside performance metrics.
Tenjerla stressed that guardrails must evolve proactively, often ahead of formal regulation, particularly as AI systems become more autonomous. Data quality and freshness, he noted, are non-negotiable, stating that currency is required for responsible decision-making. Outdated or incomplete data can compromise outcomes, erode trust and amplify risk. In this context, governance becomes a living framework—one that continuously adapts to emerging threats, regulatory expectations and operational realities.
Governance as a Built-In Foundation
Malhotra extended the conversation to the architectural level, emphasising that governance cannot be retrofitted onto agentic AI systems after deployment. “We definitely need to embed governance as part of the overall process,” he said, advocating for design principles that prioritise accountability from the outset. According to him, granular metrics, comprehensive event capture and end-to-end traceability form the backbone of responsible AI systems.
Such capabilities enable organisations to understand not only what decisions an AI system makes, but also why and how those decisions were reached. This level of transparency is critical for auditability, regulatory compliance and long-term scalability, particularly as agentic systems become more deeply integrated into core business functions.
A Clear Path Forward
The panel concluded with a shared conviction that agentic AI is not a passing trend, but a structural shift in how intelligence is operationalised across the financial sector. However, its adoption must be guided by discipline, foresight and humility. As governance evolves from policing outputs to supervising intent, organisations face a dual challenge: to innovate rapidly while maintaining trust, accountability and control.
The message from Mumbai was unequivocal. The future of agentic AI will not be defined solely by technological sophistication, but by the strength of the governance foundations that support it. In an era of autonomous action, trust must be engineered with the same rigour as the systems themselves.
Published – December 23, 2025 04:10 am IST
