The Data Security Council of India (DSCI) has released an industry guidance framework to address a growing question in artificial intelligence: who should be held responsible when an autonomous AI agent makes a mistake, causes financial damage or creates a security breach? The proposal recommends distributing accountability across the AI technology value chain, based on who controls the system, makes key decisions and has the ability to prevent harm.
Titled AI Liability: Industry Guidance for Agentic AI, the document aims to help Indian businesses understand their responsibilities when deploying increasingly autonomous AI systems. Rather than proposing a new law or regulatory standard, DSCI explains how existing legal principles, contractual obligations and sector-specific regulations can apply to agentic AI. The guidance is intended as a practical industry reference and has been opened for stakeholder feedback.
Why AI Agents Are Raising New Liability Questions
Traditional generative AI tools generally respond to user prompts by producing text, images, code or other outputs. Humans typically review those outputs before deciding what to do with them.
Agentic AI systems operate differently. They can plan tasks, access external tools, interact with application programming interfaces (APIs), retrieve information and execute instructions with limited human intervention. Some systems can also delegate tasks to other agents, creating complex workflows involving several models, applications and service providers.
This autonomy creates new risks for businesses. An AI agent connected to a company’s financial systems, for example, could initiate an incorrect payment, expose confidential information or make an unauthorised change to a business application. If multiple agents interact, an initial error could spread across connected systems.
Determining responsibility becomes difficult when the system involves several parties. The model developer may have created the underlying technology, another company may have integrated it into an application, and the deploying organisation may have configured its permissions and approved its use.
DSCI argues that assigning liability exclusively to the company using the AI or to the technology provider would overlook the shared nature of these systems.
DSCI Maps Accountability Across the AI Value Chain
The proposed guidance identifies five major groups involved in bringing AI agents into enterprise environments, while examining AI model providers separately. Each group has responsibilities linked to the decisions and controls it manages.
| Stakeholder | Areas of responsibility |
|---|---|
| End-user organisations | Selecting use cases, setting permissions, assessing risks and deciding human approval requirements |
| Service providers and system integrators | System design, configuration, integration and approval mechanisms |
| Product companies and startups | Product claims, default settings, product safety and model versions |
| Tool and framework providers | Orchestration, memory, agent coordination and safety features |
| Cloud and infrastructure providers | System isolation, identity management, data retention and platform logs |
| AI model providers | Model capabilities, safety measures, disclosures and notifications about significant changes |
The framework links accountability to the degree of control each participant exercises over an AI system.
For example, a company that gives an AI agent broad access to sensitive databases or payment systems may carry greater responsibility for the consequences of those permissions. Similarly, an integrator responsible for designing the approval process may face accountability if that process fails to prevent an unauthorised action.
This approach is intended to make responsibility more specific instead of treating AI liability as a simple choice between the developer and the user.
Open-Source and Closed AI Models May Create Different Responsibilities
The way an organisation obtains and operates an AI model can also influence its liability exposure.
Companies using open-source or open-weight models on their own infrastructure may assume greater responsibility for auditing, modifying and monitoring those models. They have more direct control over deployment choices, system configuration and operational safeguards.
Businesses accessing closed AI models through managed application programming interfaces may have contractual routes to seek recourse from model providers. However, the extent of that protection will depend on the agreements between the parties, including liability limitations and other contractual conditions.
DSCI also maintains that model providers should remain accountable for the capabilities they develop, safety measures they implement, disclosures about known limitations and notifications of significant changes in model behaviour.
The framework therefore recognises that responsibility can shift depending on who controls the model, where it is hosted and how much authority other participants have over its operation.
Twelve Categories of Potential AI-Related Harm
DSCI identifies 12 areas of potential harm associated with agentic AI systems. These include:
- Privacy breaches and unauthorised access to personal information.
- Cybersecurity threats and system vulnerabilities.
- Discrimination and unfair outcomes.
- Safety-related risks.
- Inaccurate or misleading outputs.
- Intellectual property violations.
- Deepfakes and synthetic media misuse.
- Loss of control over autonomous systems.
- AI-enabled cybercrime.
The document also highlights operational difficulties that can make investigations complicated. These include changes in model behaviour, decisions made without direct human intervention, approval fatigue among employees and the delegation of tasks between multiple agents.
Another challenge is reconstructing how an AI agent reached a particular decision. Without adequate logs and records, organisations may struggle to establish which instruction, model response, tool call or human approval contributed to an incident.
Six Safeguards to Reduce AI Agent Risks
To address these risks, DSCI recommends a layered approach to AI governance. Its guidance outlines six broad safeguards that organisations should consider when designing and deploying autonomous agents.
1. Remove unnecessary capabilities: Organisations should limit the functions available to an AI agent. An agent performing a narrow task should not automatically receive access to unrelated systems or sensitive information.
2. Restrict permissions: Access to databases, applications, financial systems and external tools should be limited to what the agent needs to perform its assigned responsibilities.
3. Require human approval: Significant, sensitive or irreversible actions should require human authorisation. This could include large financial transactions, changes to critical infrastructure or the deletion of important data.
4. Monitor agent behaviour: Organisations should continuously monitor AI activity, including tool calls, system interactions and unusual patterns that may indicate errors or misuse.
5. Enable safe fallback mechanisms: Businesses should ensure that agents can reduce their activity, return to a controlled state or stop executing tasks when unexpected behaviour occurs.
6. Establish kill switches: Organisations should maintain emergency shutdown mechanisms with clearly assigned ownership. Employees must know who can disable an agent and under what circumstances.
The guidance also recommends conducting risk assessments before deployment, including liability provisions in procurement contracts and regularly reviewing security and governance controls.
DSCI Invites Industry Feedback
DSCI has invited AI developers, enterprise leaders, government bodies, academic institutions, legal professionals and other stakeholders to examine the guidance against real-world deployments.
The council intends to use industry feedback to improve the framework and develop further guidance on areas such as authority mapping, AI deployment in physical and biological systems, and agents operated by customers and business partners.
The proposal arrives as India increases its focus on responsible AI adoption. MeitY has been working on data-sharing mechanisms for AI training, while the Telecom Regulatory Authority of India has also released draft guidelines on responsible AI use in telecom networks and services.
These initiatives reflect a broader effort to address the governance challenges created by increasingly capable AI systems without unnecessarily slowing innovation.
The Bigger Picture
The growth of agentic AI is changing the relationship between software developers, technology providers and businesses. When an AI system can independently access tools, make decisions and execute actions, accountability becomes more complicated than it is with conventional software.
For Indian enterprises, the DSCI guidance offers a way to think about liability before deploying autonomous systems at scale. It encourages companies to identify who controls each decision, establish clear approval processes and maintain records that can help explain what happened when something goes wrong.
However, the document is industry guidance, not a new binding law. Its practical influence will depend on how businesses, legal experts and policymakers use its recommendations alongside existing laws and contractual arrangements.
Looking Ahead
As companies move from AI assistants towards autonomous agents, questions about responsibility are likely to become more important. Businesses will need to evaluate not only the accuracy of AI-generated outputs but also the permissions, integrations and actions that their systems can execute independently. DSCI’s framework could help enterprises establish clearer accountability before these technologies become deeply embedded in critical operations.
The next step will be translating the guidance into practical governance policies across different industries. Its effectiveness will depend on whether organisations adopt meaningful human oversight, restrict unnecessary permissions and maintain reliable audit trails. As agentic AI adoption expands, clearly defined responsibility could become an important condition for building trust in enterprise AI systems.
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
