As Indian enterprises move AI agents from pilots into live workflows, one question keeps surfacing in boardrooms and legal teams: if an agent takes an action on its own and causes harm, who answers for it? The Data Security Council of India (DSCI) has now put forward its answer in a new discussion paper, AI Liability: Industry Guidance for Agentic AI, and its central message is that the law India needs largely exists already. What is missing is the engineering to prove that controls actually worked.

The paper, listed on DSCI’s website on 29 October and opened for industry consultation, offers what the council calls a practical, India-focused view of how responsibility should be shared across the many players that bring an AI agent into use. According to DSCI, it does not propose new standards. Instead, it shows how established legal doctrines apply to autonomous systems that are often built from several vendors’ components.

Old doctrines, new software

As reported by Express Computer, the guidance maps long-standing legal principles onto the agentic AI stack. These include causation, reasonable foreseeability, apparent authority, principal-agent law, duty of care, comparative negligence and product liability. DSCI’s view is that courts will reach for these familiar tests. The real gap lies in applying them to software that exercises its own discretion.

The paper also rejects the idea of naming a single culprit. Blame is not simply a choice between the enterprise that deploys an agent and the model or tool providers further up the chain. Its organising principle is that each party is responsible for what it was able to decide and had the tools to enforce.

Intent is not control

For CIOs and CISOs, the most useful idea in the paper may be the line it draws between what an organisation intends and what it can actually enforce. DSCI describes three layers: governance intent (what policy says), decision space (whether the architecture allows that choice at all), and instruments of execution such as permission scopes, logs, approval interfaces and circuit breakers. Liability, it argues, tends to sit wherever that chain breaks.

Take a refund limit that exists only as a line in an agent’s prompt. It is a very different thing from a limit hard-coded into the payment tool itself. The paper warns that a control which is documented but never built, or built but switched off, is worse than having no control, because it shows the risk was known and left open. It also cautions vendors: model providers and integrators that withhold the instruments customers need to enforce policy may expose themselves to liability.

More control, more ownership

The guidance notes that early architecture choices quietly redistribute risk. Enterprises that pick open-weight models or move from public cloud APIs to private or on-premises hosting gain control, but they also take on more of the responsibility when things fail. The one choice that lowers exposure in every setup, DSCI says, is strong engineering around the model: identity, delegation, memory design and guardrails.

On kill switches, the paper takes a measured line under the heading “Prevent First, Halt Last”. It sets out a graded approach. First, avoid granting risky capabilities at all. Then constrain permissions, require human confirmation only for material and hard-to-reverse actions, monitor for drift, and degrade gracefully. A scoped circuit breaker is the last resort, because stopping an agent mid-task can itself leave a payment or process half done.

DSCI CEO Vinayak Godse said the guidance aims to move AI governance “from principle to practice”, adding that intent alone is not enough and organisations need decision spaces and instruments they can actually implement. DSCI plans follow-on guidance and has invited CIOs, CISOs and risk leaders to test the paper against real deployments.

The paper arrives as the Centre prepares a consultation paper on AI regulation and debates how to treat AI platforms (see Elets CIO’s coverage of the ‘AI intermediary’ debate). For enterprises, it works as an audit checklist. For every agent in production, can the organisation show where policy becomes an enforced setting, which role it plays in the value chain, and what the agent saw, decided and did?

Sources: Data Security Council of India; Express Computer.

Be a part of Elets Collaborative Initiatives. Join Us for Upcoming Events and explore business opportunities. Like us on Facebook , connect with us on LinkedIn and follow us on Twitter.

"Exciting news! Elets technomedia is now on WhatsApp Channels Subscribe today by clicking the link and stay updated with the latest insights!" Click here!