The Data Security Council of India (DSCI) has released an industry guide on artificial intelligence accountability tailored for India, focusing on agentic AI. This document can significantly assist in determining who is responsible if an agent goes out of control or breaches its duties.
Instead of proposing new standards, the guide clarifies how existing mandatory doctrines apply to autonomous AI systems from various vendors and shows where responsibility lies throughout the value chain that brings AI into use.
It examines the use cases of agentic AI, the types of impacts and risks these systems can create, as well as the factors determining accountability. Furthermore, the document explains why a proportional and clearly distributed approach to governance, oversight, and responsibility is important for India to guide the user industry and accelerate responsible AI adoption.
The report notes that defining responsibility will be crucial depending on where the models are hosted, as well as who owns and controls them. The choice of hosting distributes responsibility differently. The architecture and hosting selection determine whether responsibility is distributed across the entire value chain or concentrated within the enterprise, which creates a central paradox of AI deployment in the corporate environment. Controlling the architecture and ownership of decision-making spaces removes an external shield of liability, and enterprises are likely to become the primary legal owners of both the infrastructure and the agent's behavior.
Responsibility essentially rests with five groups utilizing AI: user organizations, service providers and integrators, product companies and startups, tool and framework providers, cloud and infrastructure solution providers, and model providers.
The guide cautions against viewing responsibility as a binary choice between the end-user organization and participants in the value chain. It warns that framing this choice as either 'or' between the end-user and the upstream value chain participant will be legally incorrect, technically inaccurate, commercially paralyzing, and undermine a level playing field.
Moreover, the report indicates that AI permeates enterprises through various channels such as browsers, co-pilots, generative AI in enterprise software, as well as AI models and tools. Each of these pathways carries its own question of accountability, so each path requires attention and careful examination. Unknown or unidentified risks may lead to the enterprise bearing responsibility without being able to take action.
