What is Singapore’s Model AI Governance Framework?
By Nick Jang, founder of MINERVA Studio ·
The short answer
The Model AI Governance Framework is Singapore’s voluntary guidance for organisations that build or use AI, published by the Infocomm Media Development Authority (IMDA) and the Personal Data Protection Commission. There are now three: the original framework (2019, second edition 2020), the Model AI Governance Framework for Generative AI (30 May 2024, nine dimensions), and the Model AI Governance Framework for Agentic AI, launched at Davos in January 2026 and updated to version 1.5 on 20 May 2026. The agentic framework is organised around four dimensions: assess and bound the risks upfront, make humans meaningfully accountable, implement technical controls and processes, and enable end-user responsibility.
Three frameworks, one line of thinking
| Framework | Published | Organised around |
|---|---|---|
| Model AI Governance Framework | 23 January 2019; second edition 21 January 2020 | Internal governance, the level of human involvement in AI-augmented decisions, operations management, and stakeholder communication. |
| Model AI Governance Framework for Generative AI | 30 May 2024 | Nine dimensions: accountability, data, trusted development and deployment, incident reporting, testing and assurance, security, content provenance, safety and alignment R&D, AI for public good. |
| Model AI Governance Framework for Agentic AI | January 2026; version 1.5 on 20 May 2026 | Four dimensions, covered below. |
None of them is law. They’re practical guidance, written with companies and full of case studies, and they’re often quoted outside Singapore because they’re specific in a way many principle documents aren’t.
The agentic framework’s four dimensions
The 2026 framework is for organisations deploying AI agents, whether they build them or buy them. It starts from a simple point: agents can take actions, not just produce text, so access to data and the ability to change things, like updating a customer record or making a payment, cut both ways.
| Dimension | What it means in practice |
|---|---|
| 1. Assess and bound the risks upfront | Pick suitable use cases; judge risk by the scope of actions, whether they can be undone, and the level of autonomy; limit tools and access by design; give each agent its own identity and permissions. |
| 2. Make humans meaningfully accountable | Assign responsibility inside the organisation and with vendors; set checkpoints that need human approval for high-stakes or irreversible actions; check that human oversight still works, for example by watching override rates and response times. |
| 3. Implement technical controls and processes | Controls for planning, tools and protocols during development; test agents before deployment for task accuracy, policy adherence and tool use; roll out gradually and monitor continuously; manage changes so a small change can’t have an outsized effect. |
| 4. Enable end-user responsibility | Tell users what the agent can do, what data it reaches and what they’re responsible for; train staff who work alongside agents, and keep their own skills from fading. |
Version 1.5 added, among other things, systemic and multi-agent risks, new risk factors such as system complexity and third-party solutions, a split between platform providers and app developers in the agentic value chain, more practices against automation bias, and guidance on choosing between structural, rule-based and prompt-layer controls. IMDA calls it a living document.
Agent identity and approval checkpoints
Two ideas from the framework are worth lifting directly. First, an agent should have its own identity, unique and verifiable, tied to a supervising person or system, with permissions that are scoped, time-bound and follow least privilege. Without that, nobody can say afterwards which agent did what, or on whose behalf.
Second, human approval should sit at defined checkpoints, not everywhere. Irreversible actions such as deleting data, sending messages or making payments need approval; routine, reversible ones can be reviewed after the fact. Approval on everything doesn’t scale, and people start clicking yes without reading, which is the automation bias the framework keeps returning to.
What a Canadian bank can take from it
Singapore’s framework doesn’t apply in Canada, but it lines up well with what Canadian regulators expect. OSFI’s E-23 treats AI agents built on models as models to be inventoried, rated, validated and monitored; OSFI’s AI bulletin sets out practices for agents; and B-10 covers agents bought from a vendor. The agentic framework fills in the how: bound the agent’s permissions, give it an identity, put approvals where actions can’t be undone, and test and monitor it like any other model.
MINERVA does those things for agents: each agent is recorded with its tools and permissions, its policy is decided by a person and enforced on every tool call, held actions wait for a named approver, and every decision lands on a verifiable record. See agentic AI governance and Know your agent.
Related questions
Is Singapore’s Model AI Governance Framework mandatory?
No. All three frameworks are voluntary guidance from IMDA and the PDPC. Organisations use them as practical references; they are not regulations.
What are the four dimensions of Singapore’s agentic AI framework?
Assess and bound the risks upfront; make humans meaningfully accountable; implement technical controls and processes; and enable end-user responsibility.
When was the Model AI Governance Framework for Agentic AI published?
IMDA launched it at the World Economic Forum in Davos in January 2026 and published version 1.5 on 20 May 2026, incorporating feedback from more than 60 companies.
What are the nine dimensions of Singapore’s generative AI framework?
Accountability; data; trusted development and deployment; incident reporting; testing and assurance; security; content provenance; safety and alignment R&D; and AI for public good.
Does the framework apply to AI agents bought from vendors?
Yes. The agentic framework is written for organisations deploying agents whether they build them in-house or use third-party agentic solutions, and it asks for responsibilities to be set with vendors too.
Sources
- IMDA — Model AI Governance Framework for Agentic AI, version 1.5 (20 May 2026)
- IMDA — Updated Model AI Governance Framework for Agentic AI (20 May 2026)
- AI Verify Foundation — Model AI Governance Framework for Generative AI
- PDPC and IMDA — Model AI Governance Framework, second edition (2020)
New answers are posted on MINERVA on LinkedIn. Follow along, or send the question you’d like answered next to minerva@minerva.studio.
A plain-language explanation, not legal advice. The regulator’s own text is the source; read it, and take advice on how it applies to your institution.