At the 27th ET Edge CIO&Leader annual conference in Jaipur, Dr Pavan Duggal, Senior Advocate at the Supreme Court of India and Architect of Global AI Accountability, delivered a sobering keynote on what he called the “agentic AI playbook”, a survival guide for enterprise leaders navigating the collision between autonomous AI agents and India’s Digital Personal Data Protection (DPDP) Act.
The law wasn’t built for agents
Duggal’s core warning: the DPDP Act still assumes a human hand behind every data decision. AI agents, by contrast, act on a single initial command and then operate independently, subdividing tasks, processing data, and making decisions with no further human input. Consent, he noted, was given only for the original instruction, not for what the agent does next. Whether that consent can even be withdrawn mid-task remains legally unresolved, pending clarity from the Data Protection Board.
When the agent becomes the fiduciary
Under the DPDP framework, obligations fall heavily on the “data fiduciary”, the entity deciding how personal data is used. Duggal argued that in agentic AI deployments, the agent itself effectively becomes the fiduciary, a scenario the law never anticipated. Compounding this, when an AI agent hallucinates and causes harm, liability is already trending toward the enterprise that deployed it rather than the technology vendor.
CIOs and CISOs in the crosshairs
Non-compliance under DPDP can trigger penalties up to ₹250 crore, and Duggal predicted a wave of job losses among CIOs and Data Protection Officers, who he said would be the first to be held accountable when boards demand due diligence. He flagged a widening “agentic risk map” for enterprises: shadow agents deployed without security or DPO review; non-human identity sprawl; data exfiltration via AI tools; vendor-agent opacity; unmapped cross-border data flows; and gaps in audit trails.
A ten-point survival playbook
Duggal closed with concrete guidance for enterprises: treat every agent’s action as the organisation’s own act, since courts will; inventory every AI agent, including those embedded in vendor SaaS tools; tier agents by autonomy and data sensitivity; keep a human-in-the-loop gate at every consequential action; build consent scope directly into agent design; give agents least-privileged, short-lived identities; log every action immutably; install and test kill switches before they’re needed; renegotiate vendor contracts for liability and logging obligations; and rehearse for an eventual agentic AI breach.
The clock is ticking
With DPDP enforcement beginning November 13, 2026, and jurisdictions from Singapore to the broader global community moving toward agentic AI liability frameworks, Duggal urged Indian enterprises to prepare now rather than wait for legislative clarity. His message was unambiguous: the agentic AI era, he said, is not waiting for anyone’s compliance programme and the only real choice organisations have is to get onto the bandwagon, carefully, before the “bad day” arrives.