Singapore combines binding rules with published frameworks. MAS supervises financial institutions through its Technology Risk Management Guidelines, outsourcing and conduct expectations, and the FEAT principles for AI. The PDPC enforces the Personal Data Protection Act 2012 for agents' use of personal data. IMDA's Model AI Governance Framework sets out the governance practice regulators and customers expect.
What do the rules require, in plain words?
MAS. Financial institutions must manage technology risk under the TRM Guidelines: governance by the board and senior management, secure development, change management, access control, resilience and monitoring. Outsourcing expectations require due diligence, contractual protections and oversight of service providers. The FEAT principles set out fairness, ethics, accountability and transparency in the use of AI and data analytics. MAS continues to develop its supervisory expectations on AI risk management, so check for the latest material.
The PDPA 2012. Organisations must obtain consent or rely on an exception, use data only for purposes a reasonable person would consider appropriate and that were notified, keep it accurate, protect it, retain it only as long as needed, transfer it abroad only with comparable protection, and notify significant data breaches. The PDPC has issued advisory guidelines on using personal data in AI systems that make recommendations or decisions.
IMDA's Model AI Governance Framework. A voluntary framework built on two principles: AI decisions should be explainable, transparent and fair, and AI systems should be human-centric. It gives practical guidance in four areas: internal governance structures, the level of human involvement in AI-augmented decisions, operations management, and stakeholder interaction and communication.
Who does it apply to?
MAS rules apply to banks, insurers, capital markets intermediaries, payment institutions and other financial institutions it regulates. The PDPA applies to private organisations collecting, using or disclosing personal data in Singapore. IMDA's framework is open to any organisation. Whether the framework speaks to agents in particular is examined in does Singapore's Model AI Governance Framework cover AI agents.
What does this mean when the system is an agent that acts?
Singapore's frameworks share a theme: the level of human involvement should match the probability and severity of harm. The Model Framework describes three modes: a person in the loop who makes the decision, a person over the loop who monitors and can intervene, and a person out of the loop. For agents, this maps onto how far the agent may act.
Map each capability to a level of human involvement. An agent that answers product questions can run with oversight over the loop. An agent that blocks a payment or declines a claim needs a person in the loop, or very strong evidence that it performs reliably and fairly. The judgement is made per capability, not per agent.
TRM applies to the agent as to any system. Agent identities need privileged access management, changes to an agent's instructions go through change control, and the agent's dependency on an external model provider is part of resilience planning.
FEAT asks who answers for the outcome. Each agent needs an accountable owner, and decisions affecting customers need to be explainable and assessed for fairness. Our sister article on Australia, how AI agents are regulated in Australia, shows a regulator-led approach with a different emphasis.
PDPA consent and notification follow purpose. If an agent uses customer data for a new purpose, it must be notified and, unless an exception applies, consented to.
For sector detail, see AI agents for operations leaders in insurance in Singapore and AI agents for compliance teams in payments in Singapore.
How do you meet the rules?
| Requirement | Control | Evidence it leaves |
|---|---|---|
| TRM: governance | Board-approved policy covering agents; accountable owner per agent | Policy; owner register |
| TRM: access control | Unique agent identity, least privilege, privileged access controls | Access maps; reviews |
| TRM: change management | Changes to briefs, sources and limits approved and tested | Change records |
| TRM: resilience | Fall-back when an agent or its model provider is unavailable | Continuity plans; test results |
| Outsourcing | Platform due diligence, contract, ongoing oversight | Due diligence file; reviews |
| FEAT: fairness and accountability | Fairness assessment for decision capabilities; named owner | Assessment records |
| FEAT and Model Framework: transparency | Customers told when an agent is involved and how to reach a person | Disclosures; escalation logs |
| Model Framework: human involvement | Level of involvement set per capability by severity and probability of harm | Capability risk matrix |
| PDPA: consent, purpose, retention | Purposes notified; agent memory retention set | Notice; retention schedule |
| PDPA: transfer and breach | Transfer safeguards for offshore processing; breach assessment | Transfer records; breach log |
What will MAS or the PDPC ask?
- Which agents make or support decisions about customers, and who owns each?
- How did you decide the level of human involvement for each capability?
- How are changes to an agent's instructions controlled?
- What happens to the service if the model provider fails?
- Have decision capabilities been assessed for fairness, and with what result?
- Where is personal data processed, and how is it protected when it leaves Singapore?
What do these rules not cover?
Singapore has no AI-specific statute that licenses or bans agents. The Model Framework and FEAT are not laws, although MAS expects regulated firms to apply FEAT in substance. Cross-border firms must also meet the rules in their other markets; compare how AI agents are regulated in Canada. For the full requirement-to-control map, see AI agent regulation, mapped to controls.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
Questions people ask
Is a person over the loop enough for an agent that moves money?
It depends on severity and reversibility. Payments that are hard to undo usually call for a person in the loop, or tight limits with release by a person above a threshold. Document the reasoning in your capability risk matrix.
Does the PDPA have a rule specifically on automated decisions?
The PDPA does not have a provision like the EU's on solely automated decisions. The PDPC's advisory guidelines explain how consent, notification and accountability apply to AI systems that recommend or decide.
Are FEAT principles mandatory?
They are principles, not rules with penalties. MAS expects financial institutions using AI to apply them, and supervisory reviews look at whether they have.