Last updated:
Singapore's Model AI Governance Framework is voluntary guidance issued by the IMDA for deploying AI responsibly. It now exists in three editions: the original framework first published in 2019, a generative AI edition released in 2024, and a Model AI Governance Framework for Agentic AI launched on January 22, 2026.
What are the three editions of the framework?
Singapore has extended its guidance as the technology has moved. The original Model AI Governance Framework was issued by the Infocomm Media Development Authority and the Personal Data Protection Commission in 2019, with a second edition in 2020, covering internal governance, human involvement in decision-making, operations management and stakeholder communication. A generative AI edition followed in 2024, addressing hallucination, provenance, copyright, and the systemic risks that come with foundation models. On January 22, 2026, the IMDA launched a Model AI Governance Framework for Agentic AI at the World Economic Forum, aimed at systems that plan and act with limited supervision.
Why does the agentic edition matter?
The earlier editions assume a model produces an output that a person then acts on. Agentic systems collapse that gap: the system selects its own steps, calls tools and takes actions with downstream consequences. The 2026 framework addresses the controls that follow from this, including how much autonomy an agent is granted, how its actions are bounded and logged, how humans stay meaningfully in the loop, and how accountability is assigned when several agents interact. Organizations that built their AI governance around the 2019 or 2024 editions will find those controls do not cover agent behavior.
Is the framework legally binding?
No. Singapore has not enacted a general AI statute comparable to the EU AI Act, and the Model AI Governance Framework is voluntary guidance. Its influence comes from adoption, not enforcement. Sector regulators, most notably the Monetary Authority of Singapore through its FEAT principles for financial services, set their own expectations, and firms operating in regulated sectors are usually meeting both at once. Treating the framework as a floor for internal policy while mapping sector rules on top is the common approach.
How does AI Verify relate to the framework?
AI Verify is the testing and assurance side of the same program. It gives organizations a way to evidence what the framework asks for, combining process checks against governance principles with technical tests on the system itself. It is maintained as an open-source project by the AI Verify Foundation, a not-for-profit subsidiary of the IMDA. The framework describes what good governance looks like; AI Verify is how a firm demonstrates it, which matters when a customer or regulator asks for proof.
Who should use it?
The framework is written to be sector-neutral and jurisdiction-portable, so organizations outside Singapore use it as a structured checklist even where it carries no legal weight. It is most useful to firms operating across Asia-Pacific who need one governance vocabulary that maps onto multiple national regimes, and to anyone deploying agentic systems who wants a published reference point for autonomy and oversight controls.
Real world example:
A Singapore-based payments firm deploys an agent that investigates disputed transactions, pulls records from three internal systems and drafts a resolution. Under the 2019 framework it documents internal governance and the human review step. The 2026 agentic framework requires more: the firm defines the agent's permitted action set, caps the value of any adjustment it can propose without sign-off, logs every tool call for audit, and specifies what happens when the agent's confidence falls below threshold. It then evidences the controls through AI Verify process checks so that the Monetary Authority of Singapore and its enterprise customers can see the same assurance record.
