Singapore Model AI Governance Framework: From Traditional AI to Agentic AI

Overview: Three Generations of AI Governance

Singapore has pioneered AI governance in Asia through its Model AI Governance Framework, developed by the Infocomm Media Development Authority (IMDA). The framework has evolved through three generations:

1. Model AI Governance Framework (2020): Covers traditional AI systems — recommendation engines, classification models, predictive analytics
2. Model AI Governance Framework for Generative AI (2024): Addresses risks specific to LLMs, image generators, and other generative models
3. Model AI Governance Framework for Agentic AI (January 2026): The world's first governance framework for autonomous AI agents

All three are voluntary, but they represent the gold standard for AI governance in Southeast Asia and are referenced by regulators across the region.

The Original Framework (2020)

Two Guiding Principles

1. Organizations using AI should ensure that AI decision-making is explainable, transparent, and fair.

2. AI solutions should be human-centric.

Four Key Areas

Internal governance structures and measures:

• Clear roles and responsibilities for AI governance
• Risk management framework appropriate to the organization's AI maturity
• Regular review and update of AI governance practices

Determining the level of human involvement in AI-augmented decision-making:

• Risk-based approach to human oversight
• Higher-risk decisions require more human involvement
• Clear escalation procedures

Operations management:

• Data management practices
• Model development and testing processes
• Monitoring and performance tracking
• Incident management

Stakeholder interaction and communication:

• Transparency about AI use with customers and affected parties
• Channels for feedback and complaints
• Regular disclosure of AI governance practices

The GenAI Framework (January 2024)

Developed in collaboration with 70+ organizations, this framework addresses risks specific to generative AI:

Nine Focus Areas

1. Accountability: Defining who is responsible for GenAI outputs across the development-deployment chain
2. Data: Ensuring training data quality, addressing copyright, and managing data privacy
3. Trusted development and deployment: Safety testing, red-teaming, and staged rollouts
4. Incident reporting and management: Detecting, reporting, and responding to GenAI incidents
5. Testing and assurance: Evaluating model performance, safety, and reliability
6. Content provenance: Mechanisms to identify and label AI-generated content (watermarking, metadata)
7. Safety and alignment: Ensuring GenAI behaves as intended and aligns with human values
8. Cybersecurity: Protecting against adversarial attacks, prompt injection, and data poisoning
9. Human oversight of GenAI: Maintaining meaningful human control

Practical Implementation

The framework recognizes that GenAI governance involves multiple stakeholders:

• Model developers: Responsible for safety testing, documentation, and model-level safeguards
• Application developers: Responsible for appropriate use, additional safeguards, and user-facing governance
• Deployers: Responsible for organizational governance, user training, and monitoring

The Agentic AI Framework (January 2026)

Unveiled at the World Economic Forum in Davos on 22 January 2026, this is the world's first governance framework specifically for agentic AI — AI systems that can autonomously reason, plan, and take actions.

Key Concepts

What is Agentic AI? AI systems that can:

• Set and pursue goals autonomously
• Make decisions and take actions without human intervention at each step
• Interact with external tools, systems, and other AI agents
• Adapt their behavior based on outcomes

Governance Challenges Unique to Agentic AI

• Cascading actions: Autonomous agents can trigger chains of actions, making it harder to predict and control outcomes
• Multi-agent coordination: When multiple AI agents interact, governance becomes more complex
• Accountability gaps: When an agent acts autonomously, who is responsible for the outcome?
• Safety boundaries: How to define and enforce limits on autonomous agent behavior

Framework Provisions

• Clear human accountability for autonomous agent actions
• Defined boundaries and safety limits for agent behavior
• Monitoring and intervention mechanisms
• Logging and audit trails for agent decisions and actions
• Incident management for autonomous agent failures

AI Verify and ISAGO

AI Verify

Singapore's government-developed toolkit for testing AI systems:

• What it does: Provides technical testing for AI fairness, transparency, robustness, and other governance attributes
• Status: Open-source, actively maintained by the AI Verify Foundation (90+ member organizations)
• Use case: Companies can use AI Verify to demonstrate alignment with the Model Framework

Project Moonshot

The world's first open-source LLM evaluation toolkit:

• Red-teaming capabilities
• Benchmark testing
• Baseline safety evaluations for large language models

ISAGO (Implementation and Self-Assessment Guide for Organizations)

A companion tool that helps organizations evaluate their AI governance maturity:

• Self-assessment questionnaire
• Gap analysis against the Model Framework
• ISAGO 2.0 (released 2025) integrates with AI Verify for technical testing

How to Implement

Phase 1: Foundation

1. Assign AI governance ownership (could be existing risk/compliance function)
2. Complete the ISAGO self-assessment to understand your current maturity
3. Define your risk-based approach to AI governance

Phase 2: Traditional AI Governance

4. Implement the four key areas from the original framework
5. Conduct AI Verify testing for your most important AI systems
6. Establish monitoring and incident management processes

Phase 3: GenAI Governance

7. Map your GenAI usage against the nine focus areas
8. Implement content provenance mechanisms
9. Conduct Red-teaming using Project Moonshot
10. Establish GenAI-specific use policies and training

Phase 4: Agentic AI Governance (if applicable)

11. Identify any autonomous AI agents in your systems
12. Define safety boundaries and human override mechanisms
13. Implement logging and audit trails for agent actions
14. Establish monitoring for cascading action risks

Related Regulations

• Singapore PDPA: Mandatory data protection requirements that underpin all AI governance
• Singapore MAS AI Guidelines: Sector-specific mandatory framework for financial institutions
• ASEAN AI Governance Guide: Regional framework built on Singapore's approach
• EU AI Act: Global reference with similar risk-based governance approach