Singapore Slings Agentic AI with a Model Governance Framework: Lessons for India

Generative AI systems respond to prompts. Agentic AI systems act on them. The distinction is not merely technical, it is regulatory. An agentic AI system can independently plan multi-step workflows, execute transactions on external databases, adapt its approach based on new information, and interact with other autonomous agents to complete tasks on behalf of humans, often without waiting for human approval at each step. [1] This capacity for independent action introduces governance risks that existing AI frameworks, designed for systems that generate outputs rather than take actions, were never built to address.

The speed at which agentic AI is entering enterprise workflows is significant. Gartner forecasts that 40% of enterprise applications will incorporate task-specific AI agents by the end of 2026, up from 5% in 2025. [2] Yet only 21% of organisations report having a mature governance model for agentic AI. [3]  This gap between deployment velocity and governance readiness is precisely what Singapore’s new framework seeks to close, and what makes it relevant far beyond Singapore’s borders.

On 22 January 2026, Singapore’s Minister for Communications and Information, Josephine Teo, launched the Model AI Governance Framework ("MGF") for Agentic AI at the World Economic Forum in Davos. [4]  The MGF was developed by the Infocomm Media Development Authority ("IMDA") in collaboration with over 60 companies and builds on Singapore's earlier Model AI Governance Framework (2019, updated 2020) [5]  and the Generative AI Governance Framework (2024) [6]  . Further, the Singapore authorities published the Personal Data Protection Commission’s Proposed Advisory Guidelines on Use of Personal Data in Generative AI, on 2 June 2026, with public consultation closing on 14 July 2026.

Version 1.5 of the MGF, published on 20 May 2026 and updated on 5 June 2026, incorporates expanded guidance on speed-of-decision risks, oversight mechanisms for real-time autonomous actions, systemic and multi-agent risks, and change management processes to prevent small changes from causing outsized impacts in complex systems. [7]  The MGF is voluntary and does not create binding legal obligations. However, its cross-border implications for Indian enterprises, particularly those operating in or servicing clients across ASEAN, are immediate and practical. Alongside the updated framework, IMDA published a separate Discussion Paper on Legal Responsibility for AI Agents, acknowledging that the liability question remains unsettled and inviting public feedback (“Discussion Paper”). [8]

The MGF is structured around four governance dimensions, each carrying specific compliance takeaways for organisations deploying agentic AI.

Assess and bound the risks upfront

Before deploying an agent, organisations must evaluate the risk landscape across factors such as domain sensitivity, the breadth of actions the agent can take, whether its actions are reversible, and the degree of autonomy granted. [9]  Risk containment occurs at the design stage through least-privilege access policies, hard-coded (not merely prompt-layer) behavioural constraints, and cryptographically verifiable agent identities that are centrally tracked. The MGF illustrates this through Dayos, a Singapore enterprise AI company that classified IT support tickets into three tiers, viz., 60% fully automated, 30% requiring human sign-off, and 10% excluded from agent action entirely. [10]

Make humans meaningfully accountable

The MGF places primary accountability on the deploying organisation and demands that human oversight be substantive rather than merely ceremonial. Organisations must audit override rates and response times to detect rubber-stamping behaviour. Tencent's “CodeBuddy” agentic AI coding assistant demonstrates this principle: Read operations (reading files, directory listings) proceed without approval, but file edits, shell commands, network requests, and external tool invocations require explicit human approval. [11]

Implement technical controls and processes

During development, structural safeguards must take precedence over prompt-layer instructions. Safeguards such as the Model Context Protocol (MCP) have been developed not merely as a connectivity tool but as a governance layer that logs interactions and filters sensitive data. Pre-deployment testing must cover tool-calling accuracy, policy compliance, and emergent behaviours in multi-agent configurations. Post-deployment, the MGF recommends gradual rollouts with layered alert thresholds and immutable audit logs. [12]  Terminal 3's payroll agent illustrates this approach, where all sensitive employee data is held exclusively within a hardware-level trusted execution environment, and the agent's every action is verified against pre-credentialed parameters at hardware speed. [13]

Enable end-user responsibility

The MGF distinguishes between users who interact with agents (who need transparency about capabilities and escalation paths) and users who integrate agents into daily workflows (who need training on failure modes, hallucination patterns, and skill-degradation risks). [14]  This latter concern is novel because as agents absorb and execute entry-level tasks, professionals risk losing the foundational skills needed to perform those tasks manually during system failures.

India has made significant strides in AI governance between 2025 and 2026. The IndiaAI Safety Institute (“AISI”) was announced in January 2025 and established by the Ministry of Electronics and Information Technology (“MeitY”) under the "Safe and Trusted AI" pillar of the IndiaAI Mission. The Institute, incubated by IndiaAI Mission will be set up on a hub and spoke model with various research and academic institutions, private sector partners joining the hub and taking up projects. The AISI’s mandate includes evaluating frontier AI systems. The India AI Governance Guidelines, published in February 2026 ahead of the India AI Impact Summit, established broad principles for responsible AI development (“India Governance Guidelines”). [15]  The Office of the Principal Scientific Adviser released a Techno-Legal Framework white paper in January 2026. [16]  The Digital Personal Data Protection Act, 2023 ("DPDPA") [17]  provides a baseline for data governance, and the Supreme Court's Draft Regulations for the Use of AI in Courts, circulated in June 2026, signal judicial awareness of AI's expanding footprint. [18]

However, none of these instruments address agentic AI specifically. India's governance architecture remains oriented towards AI systems that generate outputs for human review, not systems that autonomously take actions with real-world consequences. This gap is not merely theoretical, it has immediate commercial implications for Indian enterprises operating cross-border.

Binding force

Singapore's MGF is voluntary, the EU AI Act (which becomes fully effective on 2 August 2026 barring Article 6(1) [19] ) is binding with significant penalties [20]  , and India Governance Guidelines remain advisory with no statutory enforcement mechanism.

Risk classification

The MGF uses a multi-factor risk assessment calibrated to agentic capabilities (autonomy level, action reversibility, system complexity). The EU AI Act classifies AI systems into prohibited, high-risk, limited-risk, and minimal-risk categories, but was not designed with agentic autonomy as a primary variable. India has not yet adopted a formal risk classification methodology for AI systems.

Liability allocation

The MGF allocates accountability to the deploying organisation but explicitly defers the legal liability question to its separate Discussion Paper. The EU AI Act assigns obligations based on role (provider, deployer, importer, distributor) with strict liability proposed for high-risk systems. India has no framework, statutory or otherwise, that addresses liability for harm caused by autonomous AI agents acting without real-time human oversight. 

This is the most commercially significant question for cross-border AI deployments. When an autonomous agent, operating at machine speed across the value chain of developer, platform provider, deployer, orchestrator, distributor and end-user, causes harm to a third party, existing legal frameworks struggle to assign responsibility.

The Discussion Paper advances two concepts of relevance. First, the "residual risk doctrine", i.e., reasonable mitigation, not elimination of all risk, is the appropriate standard. Second, a form of the nemo dat principle applied to agents, i.e., an AI agent cannot be granted authority greater than that possessed by the human who authorised it, and all delegations must be recorded. [21]

Indian law currently relies on analogues that were not designed for this scenario. Vicarious liability under the Indian Contract Act assumes a principal who directs an agent's actions [22] , is liable, but does not categorically accommodate an autonomous system / agent that determines its own course of action. Section 43A of the Information Technology Act, 2000 imposes liability on a body corporate for negligence in implementing reasonable security practices, but its application to AI-caused harm remains untested. General tort law principles of “reasonable care” apply, but the standard of "reasonableness" for autonomous systems has not been judicially tested or defined as of today. 

The question extends further along the value chain than existing doctrine anticipates. Liability attaches not only to deployers and end-users but potentially to developers themselves, even for autonomous agent actions they did not directly control. The Indian courts have not yet confronted a case in which the principal is an AI developer and the agent is an autonomous system, but Sections 182 and 226 of the Indian Contract Act may not be inherently incompatible with that analysis. Section 226 provides that acts done by an agent on behalf of a principal bind the principal as if done by the principal personally. Where a developer designs, trains, and deploys an agentic system that operates within parameters the developer defines, a court may find that the developer is the true principal whose authority the system exercises. The agent, in the classical contractual sense, acts for the developer even when operating autonomously within those parameters. This reading would extend vicarious liability to developers who have granted their systems the authority to act, regardless of whether any human reviewer approved each individual action. Indian enterprises and their counsel could therefore potentially treat developer-level accountability as a live risk to be managed through guardrails such as clear contractual allocation, robust indemnities, and precise definitions of what authority the developer has conferred upon the deployed system. We explored these themes in an article in 2023 [23] , where the authors concluded that “The authors believe that the established construct of ownership by a "person" will continue to be necessary for some time to enable the development of proper systems to exploit the potential of AI. Such an ownership construct is also necessary to attenuate the consequences of AI by identifying such "person/s" as liable and accountable for the actions of AI. As controls and technology develop to the point where AI can be, for lack of a better word, policed or governed as persons are, a new model will develop which will de-link the person from the intelligence, and thereby usher in a new phase in the evolution of the principles of ownership, rights, responsibilities, powers and duties.”

However, global judicial developments are beginning to fill this vacuum. In Mobley v. Workday (N.D. Cal., June 2026), a federal judge ruled that Workday, as an AI vendor (not the employer), can face liability under California civil rights law for discriminatory AI-powered hiring tools, even for plaintiffs outside California. [24]  In Nippon Life v. OpenAI (filed March 2026), the plaintiff alleges developer liability where ChatGPT effectively practised law by helping a claimant draft 44 post-settlement filings, including a fabricated case citation. [25]  A May 2026 decision from the United States District Court for the Southern District of New York, American Council of Learned Societies v. NEH, established that an organisation cannot escape liability simply by attributing the harm to its AI system. [26]  This is similar to the stance taken by the MGF under its second pillar “Make Humans Meaningfully Accountable”. These precedents signal a clear direction that courts are not waiting for legislatures. Liability is attaching to developers, vendors, organisations deploying AI as their “chosen instruments”, and not merely to end-users.

For Indian enterprises deploying agentic AI or building agentic tools for international clients, the absence of domestic regulation does not equate to the absence of legal risk.

In the Indian banking and financial services sector (“BFSI”), Indian institutions deploying agentic AI for wealth analysis, compliance automation, or customer service will need governance that satisfies both the Reserve Bank of India's (“RBI”) expectations and international best practice. The MGF's case study on OCBC's Source of Wealth analysis agent, which processes sensitive financial documents through a multi-agent pipeline with human checkpoints, offers a practical template. [27]  RBI‑regulated banks / NBFCs could consider maintaining AI risk registers, audit trails, and explainability protocols for autonomous decision systems.

The RBI's draft Guidance on Regulatory Principles for Model Risk Management, published on 24 June 2026, represents a significant and prescriptive step in this direction. [28]  The draft sets out principles-based expectations for regulated entities to establish board-approved governance frameworks for all AI and machine learning models, implement kill switches and circuit-breaker controls for models operating autonomously, maintain continuous human oversight of model outputs, and conduct systematic bias testing and performance monitoring. This is a materially different posture from the MGF's voluntary framework. The RBI model is domestic, principles-based, and supervisory in character, with the potential to harden into binding obligations once finalised. Indian financial institutions will therefore likely operate in a multi-layered environment. The MGF sets international best practice and shapes contractual expectations from Singapore-based clients and counterparties, while RBI guidance (even in draft form) signals the regulator's expectations and is likely to inform supervisory review. Compliance with only one of these frameworks is unlikely to be sufficient. Institutions would have to design their AI governance architecture to satisfy the higher standard where the two frameworks overlap, and to document clearly where they apply each framework to distinct aspects of their agentic AI deployments. 

In legal technology, platforms deploying AI agents for contract review, due diligence, or litigation support face the Heppner risk. In United States v. Heppner (S.D.N.Y., February 2026), the court held that documents generated using a consumer AI platform were neither privileged nor protected as work product. [29]  Indian legal-tech deployers must design agent workflows that preserve privilege and maintain audit trails sufficient to demonstrate human oversight of legal judgment. The threshold design principle required appears to be one that separates the provision of legal information from the unauthorised practice of law i.e. where guardrails on what a system is built to do and what it is built to refuse, needs to be baked in by legal tech developers. [30]

In healthcare, AI agents that interact with patient data or make recommendations affecting clinical decisions face regulatory scrutiny under both the DPDPA (for personal data processing) and sector-specific guidelines. 

Perhaps most significantly, India's 1,600-plus Global Capability Centres ("GCCs") servicing clients in Singapore and ASEAN will increasingly face contractual obligations to demonstrate compliance with the MGF or equivalent governance maturity. [31]  This creates a de facto extraterritorial effect, where even absent Indian regulation, commercial necessity will drive adoption.

India's regulatory trajectory should be proactive rather than reactive. Rather than waiting for an agentic AI system to cause harm and then legislating in response, India should consider developing agentic AI-specific governance guidance that accounts for the unique risks these systems pose.

The directions could draw from emerging international convergence. The appropriate model for India would not be a choice between a single universal framework and complete deference to sector-specific regulators. That framing itself presents a false dilemma. The recommended approach is a federated architecture in which a central baseline framework, established by MeitY or the AISI, sets the minimum principles applicable to all agentic AI deployments including risk categorisation, accountability allocation, mandatory audit trails for high-risk agent actions, and minimum human oversight standards calibrated to autonomy levels. Sectoral regulators such as the RBI, SEBI, and the Insurance Regulatory and Development Authority of India could then layer domain-specific requirements on top of that baseline, calibrated to the particular risks, systemic importance, and regulatory traditions of their sectors. This federated architecture would avoid two failure modes (i) the over-prescription of a monolithic universal framework that cannot accommodate the diversity of AI applications across sectors, and (ii) the fragmentation of a purely sectoral approach that creates inconsistent standards, regulatory arbitrage, and compliance confusion for enterprises operating across multiple sectors. The RBI's draft model risk management guidance is a sound example of the sectoral layer in this architecture. What India currently lacks is the baseline layer on which such sectoral guidance can build. 

India could consider three key steps. First, MeitY or the AISI should issue agentic AI-specific guidance, potentially as an addendum to the India Governance Guidelines which addresses the areas we mentioned in the preceding paragraph. Second, sector regulators should incorporate agentic AI considerations into existing frameworks. The RBI's Outsourcing Guidelines, for instance, should address scenarios where an AI agent (rather than a human vendor) is performing delegated functions. Third, the Bureau of Indian Standards should develop testing and certification standards for agentic AI systems, aligned with international benchmarks to facilitate cross-border interoperability. 

Any governance framework that India adopts should be designed using the mantra of clarity, agility, flexibility and adaptability. A framework that is clear, moves fast and easily, bends when necessary (without breaking) and changes to fit new paradigms. The agentic AI landscape is dynamic and evolving rapidly. Capabilities that are experimental today become deployed at scale within months, and deployment patterns that are novel today become standard practice within a year. Singapore's own MGF illustrates this fast-paced evolution. The framework launched in January 2026 and was already at Version 1.5 by May 2026, with meaningful substantive updates on oversight mechanisms, systemic risk, and change management. A governance framework that is rigid at inception will become obsolete before it achieves broad adoption. India should therefore build in formal version-controlled review cycles, stakeholder feedback mechanisms, and a clear process for incorporating developments in international standards, judicial interpretation, and technology capability. Frameworks that fail to evolve will constantly fall behind and will likely fail to be able to accelerate to match pace. 

The window to shape India's agentic AI governance is open. Singapore's framework, while designed for a different jurisdiction and context, provides both the vocabulary and the architecture for India to build upon. Indian enterprises, lawyers, and policymakers should engage now, before the framework hardens into a de facto international standard without Indian input. Back in 2023, we concluded [32]  that “Whilst the authors don't think that the "day" to build the new model has come, they do feel that it will be here sooner rather than later.”. Sooner is here now, as is the audible clarion call to evolve regulation with respect to agentic AI.