When an AI agent acts on its own and something breaks, who is responsible? A new discussion paper from the Data Security Council of India (DSCI), AI Liability: Industry Guidance for Agentic AI, addresses that question. Its answer is clear: no new laws are needed. Existing legal ideas need to be applied carefully to a new kind of software.
Why this is different
For decades, enterprise technology ran on a simple equation: a company set the purpose, and software carried it out in a fixed, predictable way. AI changes this. The report says large models add a third element, “planetary wisdom,” meaning general, probabilistic intelligence built for the world, not for any one bank or hospital.
The risk grows with agents. Earlier chatbots only gave answers, and a human stood between the output and any action. Agents can call tools, move money, and delete records, so that buffer disappears.
It’s not either/or
The paper rejects the idea that blame falls either on the company using the AI or on the vendor. It argues that liability attaches to specific steps in a chain, and each step has a different owner. It focuses on five groups: end-user organizations, integrators, product companies, tool and framework providers, and cloud providers. Model makers get a separate chapter.
The guiding rule is that each party answers for what it could decide and had the tools to enforce. Courts, it says, will reach for old tests such as causation, foreseeability, duty of care and shared negligence.
Your choices move the risk
The report makes a notable point about control. Using a closed model via a public API spreads liability among vendors, but you have little control. Self-hosting an open model gives you full control, and also full ownership of its failures. The more freedom you have to inspect, host, or modify the intelligence, the more exclusively you own its failures.
From intent to control
Good policy is not enough. A rule that no agent may approve refunds above Rs 25,000 counts only if the platform lets you set that limit and someone enforces it. Investigators, the paper says, will not ask what a company meant to happen, but whether it could have made it happen and whether it did. Vendors that withhold such controls also expose themselves.
Prevent first, halt last
The report also cautions against leaning on a “kill switch.” Its ladder runs from designing risk out, to limiting permissions, to human approval for big and hard-to-reverse actions, then monitoring, graceful slowdown and, only last, a halt. Stopping mid-transaction can leave a payment half-done, and the harm is still yours. A control that exists only on paper shows the risk was known.
Hidden doors for AI
The paper lists eleven ways AI enters a company, including copilots switched on by default, AI embedded in software, and customers’ own agents. It warns that courts will ask what an organization knew or ought to have known, and an unexamined route is unlikely to be treated as unknown.
What to take away
For CISOs, risk heads and lawyers, the message is practical:
- Turn policy into enforced settings.
- Keep records of what agents saw, decided and did.
- Match controls to the risk, since over-control creates its own liability.
DSCI calls the paper a living document and invites industry feedback at policy@dsci.in. Further guidance on authority maps and execution controls is planned.
