TLDR: Singapore’s Infocomm Media Development Authority has rolled out a Model AI Governance Framework for Agentic AI, first unveiled in January and updated again in May after feedback from more than 60 organizations. It’s not a law with penalties attached, but it signals where enforcement and procurement standards are headed. Companies building or deploying AI agents in Singapore should start lining up their practices now.
A Framework Born at Davos
Singapore didn’t quietly publish this one in a government gazette. Minister for Digital Development and Information Josephine Teo unveiled the Model AI Governance Framework for Agentic AI on January 22 at the World Economic Forum in Davos. That venue choice wasn’t an accident. Singapore wanted the world watching, and it got its wish: the framework is being described as the first of its kind anywhere to specifically address agentic AI, the category of systems that can take multi-step actions on their own rather than just answering a prompt.
The Infocomm Media Development Authority, known locally as IMDA, built the framework. It’s non-binding for now, which means there’s no fine for ignoring it. But in Singapore, “non-binding” guidance has a habit of becoming the baseline regulators expect anyway, especially once it shows up in government tenders and licensing conversations.
Why Agentic AI Needed Its Own Rules
General AI governance frameworks tend to focus on outputs: is the content accurate, is it biased, did it leak someone’s data. Agentic systems raise a different problem. They act. An AI agent might book a trip, move money between accounts, or adjust a factory’s production schedule without a human clicking approve at each step. IMDA’s framework maps that risk across categories including unauthorized actions, data breaches, and disruption to systems the agent is connected to.
That’s a meaningful shift for any business running automation on top of a large language model. If your chatbot used to just answer customer questions and now reconciles refunds or edits records, you’ve crossed into the territory this framework is built for.
The Four Pillars Businesses Need to Actually Implement
The framework rests on four ideas, and they’re worth treating as a checklist rather than background reading.
Assess and Bound the Risk Upfront
Before deployment, teams are expected to map out what the agent can actually do and set hard limits on it. Not a general risk statement buried in a policy document, but specific boundaries: what systems can it touch, what actions require a human sign-off first.
Keep a Human Meaningfully Accountable
IMDA’s language here is deliberate. Accountability has to be meaningful, not a name on an org chart who never looks at what the system does. Someone has to actually own the agent’s behavior and be positioned to intervene.
Build in Technical Controls and End-User Responsibility
The remaining two pillars cover the engineering side (logging, kill switches, monitoring) and the user side (making sure the people interacting with an agent understand its limits and know when to stop trusting it blindly).
What Changed Between January and May
The original framework drew feedback from over 60 organizations in its first few months, and IMDA published an updated Version 1.5 on May 20. The revision added a new section on multi-agent systemic risk, which matters as companies start chaining several agents together rather than running one in isolation. It also got more specific about technical controls and added real-world case studies, which is usually a sign regulators are preparing for wider adoption rather than treating something as a passing trend.
Where This Fits in Singapore’s Bigger AI Push
This framework didn’t appear out of nowhere. In February, Prime Minister Lawrence Wong announced a new National AI Council that he chairs personally, overseeing national AI Missions in advanced manufacturing, connectivity, finance, and healthcare. The government has also committed roughly SG$1 billion toward public AI research through 2030. Businesses can also run their systems through AI Verify, a testing toolkit IMDA offers to check AI governance practices against recognized international standards.
Put together, the signal is pretty clear. Singapore wants to be the jurisdiction companies point to when they need to show regulators elsewhere that their AI governance is credible, and it’s building the infrastructure, funding, and now the frameworks to back that up.
What to Do Before Your Next Audit
Legal and compliance teams working with clients operating in Singapore should map existing AI deployments against the four pillars now, not after a regulator or a client asks for it. Document what each agent can do, who’s accountable for it, and what happens when it fails. None of this requires a court order to take seriously. By the time it does, the companies that acted early will be the ones with paperwork ready instead of a scramble on their hands.
What businesses in Singapore are learning fastest is that the agents doing the most for them are also the ones that need the tightest leash, and the law is starting to catch up to that reality one framework update at a time.
