Navigating the Agentic Frontier: Operationalizing Responsibility Across the AI Value Chain

Grounding AI governance in enterprise realities from context engineering to non- delegable legal duties.

Historically, enterprise IT liability rested on a predictable equilibrium: Purpose (defined by the enterprise) + Means (built by deterministic software providers). Because system workflows were fixed and rule-bound, foreseeability and accountability were relatively straightforward to trace.

The introduction of Planetary Intelligence and multi-agent autonomous execution breaks this traditional balance. Agentic AI introduces probabilistic reasoning into deterministic business processes. As agents sub-delegate tasks, invoke third-party tools via protocols like MCP, and process unstructured inputs, the traditional causal buffer between model outputs and real-world actions disappears.

The Data Security Council of India (DSCI) is leading an initiative to establish a proportionate, actionable framework for Agentic AI Liability.

Key Framework Insights in the Discussion Paper:

  • The 12 Domains of Loss & Harm: Categorizing operational impact across privacy exposure, prompt injection, containment escape, deepfakes, and market-wide systemic agent risks. 
  • Impact of Deployment & Model Choice: How open-weight, open-source, self-hosted, and closed API models fundamentally alter an enterprise's legal burden of proof and recourse capabilities. 
  • The 8 Pillars of AI Engineering: Why context curation, memory/state engineering, identity delegation, and telemetry serve as the primary source of evidence in post-incident investigations. 
  • Graded Execution & Interventions: Moving away from blanket kill-switches toward a 6-step hierarchy: Design → Constrain → Confirm → Observe → Degrade Gracefully → Halt. 

📥 Open Consultation & Call for Expert Inputs


This discussion paper is designed as a living document to support enterprise leaders, risk officers, policy architects, and technologists. Before finalizing our submission, we are conducting a broad consultation to ensure the framework reflects operational reality. 

We invite you to test this framework against your organization's AI deployments and share your critical insights: 

  1. Are the decision spaces and authority boundaries between integrators, software vendors, and end-users clearly delineated? 
  1. What additional nuances should be considered for inbound agents operating across organizational boundaries? 

👉 Download the Discussion Paper: Link to PDF on DSCI website.

💬 Submit Your Inputs: Submit your detailed feedback directly through the research@iidalabs.org / policy@dsci.in.

Image Source - Lone Thomasky & Bits&Bäume / https://betterimagesofai.org / https://creativecommons.org/licenses/by/4.0/