- Explainer
- 4 min read
Singapore’s governance framework for agentic AI, explained
What the IMDA framework asks of organisations deploying AI agents, how it relates to the MAS guidelines, and where the testing kits fit.
In short
Singapore launched the Model AI Governance Framework for Agentic AI on 22 January 2026. It is guidance built around four dimensions: bound the risk upfront, keep people meaningfully accountable, apply technical controls through the agent life cycle, and help end users act responsibly. Banks and insurers also have the MAS AI risk guidelines from October 2026.
What it is
IMDA developed the framework and Minister Josephine Teo announced it at the World Economic Forum in Davos on 22 January 2026.1 It extends Singapore’s earlier Model AI Governance Framework to systems that plan and act on their own, and commentators have described it as the first governance framework written specifically for AI agents.2 It recommends measures, and IMDA calls it a living document that will change with feedback and case studies.1
The four dimensions
- Assess and bound the risks upfront. Choose suitable use cases and limit the agent’s autonomy and its access to tools and data.1
- Make humans meaningfully accountable. Set checkpoints where a person must approve.1
- Apply technical controls and processes across the agent life cycle. For example baseline testing, and access limited to whitelisted services.1
- Enable end-user responsibility. Through transparency, education and training.1
For a customer service agent, the first dimension usually means starting with read-only access, the second means approval for refunds or account changes above a limit, and the fourth means telling customers they are talking to an AI and how to reach a person.
Where the MAS guidelines come in
On 7 October 2026 MAS issued Guidelines on AI Risk Management that apply to all financial institutions and all forms of AI, including agentic systems that operate autonomously and access tools.3 They expect board and senior management oversight, an inventory of AI use, an assessment of risk materiality, and proportionate controls across the life cycle. Institutions stay accountable for AI from third parties. The guidelines take effect on 7 October 2027, and MAS plans to consult on further agentic AI guidance in 2027.3
Testing the agent
IMDA has published two starter kits. The kit for LLM-based applications, version 1.0 dated January 2026, covers pre-deployment testing for baseline safety and reliability.4 The kit for agentic systems followed in October 2026 as a practical reference for testing agents for known risks.5 Our guide to testing conversational agents before launch shows how to apply them in a service setting.
Questions people ask
Is the framework mandatory?
It is guidance that recommends measures. Financial institutions should also read the MAS guidelines, which set supervisory expectations from 7 October 2027.3
Does it apply to a chatbot that only answers questions?
It is written for agents that plan and act. A chatbot that only answers questions carries fewer of the risks it addresses, though the transparency points still apply.
Can we give feedback?
Yes. IMDA invites feedback and case studies through a form linked from the launch announcement.1
Sources
- Singapore Launches New Model AI Governance Framework for Agentic AIMinistry of Digital Development and Information, 22 January 2026
- Singapore: Governance framework for agentic AI launchedBaker McKenzie, January 2026
- MAS Sets Out Supervisory Expectations on Responsible AI Adoption by Financial InstitutionsMonetary Authority of Singapore, 7 October 2026
- Starter Kit for Testing LLM-Based Applications for Safety and Reliability, version 1.0IMDA, January 2026
- Starter Kit for Testing Agentic Systems for Safety and Reliability, version 1.0IMDA, October 2026
Keep reading
More from Resources
Conversational AI and CX events in Singapore, 2027
The main 2027 events in Singapore for conversational AI, customer experience and enterprise AI teams, with dates and venues from each organiser.
WorksheetTen questions worth bringing to CAIS 2027
Ten questions for delegates, grouped by the sessions on the agenda, with what a useful answer sounds like.
GuideTesting a conversational agent before it meets customers
How to build a test set from real conversations, test actions safely, and keep testing after launch, using Singapore’s testing kits as the frame.
CAIS 2027, Singapore
Hear it first-hand from the people doing the work
A day of keynotes, panels and 1-on-1 meetings with senior AI, customer experience and technology leaders, on 7 July 2027.