Building Responsible AI at Scale: IBM’s Blueprint for the Agentic Era

As Indian enterprises move rapidly from AI pilots to production-scale agentic systems, the conversation around Responsible AI is shifting. What was once viewed as an ethical checkbox is now a core business and regulatory requirement.

AI agents today do more than automate tasks; they fetch data, interact with systems, trigger workflows, and influence decisions that carry financial, legal, and operational consequences. In this environment, accountability, transparency, and human oversight are no longer optional design choices.

To understand how enterprises can embed trust into autonomous systems without slowing innovation, CiOL spoke with Vishal Chahal, Vice President, IBM India Software Labs, on how IBM is operationalising Responsible AI for the agentic era.

Interview Excerpts:

As enterprises shift from pilot projects to full-fledged agentic AI systems, how will accountability and liability be structured when AI-driven decisions lead to real-world legal, financial or safety outcomes? Who bears ultimate responsibility: the model builder, the integrator, or the deploying enterprise?

Accountability has to be proportionate to the risk profile of the solution. Responsible AI must be embedded at the design stage itself.

As a platform provider, we build the foundational capabilities. But creators, solution designers, and enterprises deploying these agents all have defined responsibilities. They provide the context, memory, and governance inputs that shape agent behaviour.

AI agents are autonomous in how they execute actions, not in owning the business process. Critical decisions remain human-driven, with humans always in the loop. Responsibility is distributed based on who did what.

Translating Responsible AI principles into live systems often involves engineering trade-offs. What are the most challenging trade-offs you face between model performance, explainability, compliance, and latency, especially in mission-critical enterprise applications?

The priority is always explainability, compliance, and governance. Performance matters, but never at the cost of trust.

What we increasingly see is AI becoming more specific. Enterprises don’t need the full power of a large language model for every task. In many cases, a smaller, domain-trained model works better; it’s faster, more controllable, and easier to govern.

If there’s ever a trade-off, we stay aligned with the pillars of Responsible AI. Specialisation helps us meet performance needs without compromising governance.

Human-in-the-loop oversight is often promised as a compliance safeguard, but at scale, human review may become symbolic. Which governance models do you see as genuinely effective at preserving meaningful human accountability in high-frequency or real-time AI workflows?

Agents primarily automate the IT and data-fetching side. Business decisions still sit firmly with humans.

Platforms should actively assist human decision-making, flagging bias, explaining outputs, and surfacing risk indicators. Oversight works when systems are designed to support humans, not bypass them.

⁠With agentic AI systems increasingly capable of interacting autonomously across business data, what emerging risk vectors (hallucinations, adversarial inputs, cascading agent-to-agent errors, feedback loops) concern you the most, and how are you stress-testing systems to surface these before deployment?

Unmanaged model behaviour is a real concern. That’s why governance must sit at the agent layer, not just the model layer.

In IBM’s Watson Orchestrate, agents operate under governance policies. Even if a model hallucinates, the agent won’t act unless the output complies with defined rules.

We also stress-test our systems internally through what we call ‘Client Zero’ deployments, using our own AI at enterprise scale before customers do.

For enterprises operating under strict regulation (banking, healthcare, government), how do you ensure compliance in AI-driven decision-making: through audit trails, immutable logging, real-time monitoring, or by design constraints? Which approach do you consider most robust for long-term governance?

Governance has to be embedded by design. An orchestrator with built-in compliance ensures that every agent, whether built internally or by a third party, operates within policy boundaries.

This approach gives enterprises confidence that non-compliant behaviour simply won’t execute.

⁠Bias, fairness and ethical risk aren’t just technical issues — they can lead to reputational and legal consequences. What concrete mechanisms (data lineage, impact audits, adversarial testing, red-teaming) does IBM embed to ensure fairness and accountability across geography, language and demographic variance?

Governance must exist at all layers: data, model, and agentic systems. Data lineage and traceability are foundational.

Enterprises need visibility into what data models are trained on. Without that, fairness guarantees fall apart. That’s why auditability and lineage remain non-negotiable in enterprise AI.”

Responsible AI is here to stay. It will become embedded in how AI is designed, deployed, and taught.

We’re already seeing universities and governments include Responsible AI in curricula and policy discussions. As AI scales further, especially into regulated industries, governance will increasingly show up as a default requirement in RFPs.