Singapore's Model AI Governance Framework from IMDA is voluntary guidance that fits AI agents well. It asks for clear internal governance with named owners, a level of human involvement matched to the harm a decision can cause, disciplined management of data and models in operation, and open communication with the people affected. Human involvement is the hard part for agents.
What does the framework set out?
The framework rests on two guiding principles: decisions made by or with AI should be explainable, transparent and fair, and AI systems should be human-centric, protecting people's interests, safety and wellbeing.
It turns those principles into practice across four areas, in substance:
- Internal governance structures and measures. Clear roles and responsibilities for AI, risk management and internal controls, and staff trained to carry out their part.
- Determining the level of human involvement in AI-augmented decision-making. Organisations weigh the probability and severity of harm to decide whether a person should be in the loop (approving), over the loop (supervising and able to intervene) or out of the loop.
- Operations management. Data quality and lineage, minimising bias, explainability, repeatability, robustness, regular tuning, traceability and auditability.
- Stakeholder interaction and communication. Telling people when AI is used, explaining how it affects them, giving them a way to give feedback and seek review, and testing the communications.
IMDA versions the framework and adds companion material as AI changes, so check the current set for anything agent-specific. Our short answer to whether Singapore's Model AI Governance Framework covers AI agents covers that scope question.
Who does it apply to?
Any organisation that develops or deploys AI, in Singapore or elsewhere. It is not law. Financial institutions supervised by MAS face the FEAT principles and technology risk rules alongside it, and the PDPC uses it as a reference for responsible AI practice.
What does it mean when the system is an agent that acts?
The framework does not prescribe how to build agents, and it should not be confused with a software agent framework, which is a developer toolkit. It is a governance framework. Applied to agents, each area gains a specific shape.
Human involvement must be set per action, not per agent. The probability-severity matrix in the framework works well for agents if you apply it to each thing the agent can do. Reading a policy document is low severity; a person can stay out of the loop. Drafting a claim decision is higher severity; a person should be over or in the loop. Paying out a claim is highest; a person in the loop until the agent has proven itself, and some actions perhaps never.
Operations management covers instructions and sources, not only models. For agents, traceability means recording which facts the agent used, from where, and which limits applied. Repeatability means testing the agent's behaviour on the same cases after each change to its brief or sources.
Model choice is a governance decision. An agent tied to one model provider inherits that provider's changes. Being able to change model without losing governance evidence is part of operations management; our article on why model independence is a governance requirement makes that case.
Communication must reach the moment of interaction. People should know when they are dealing with an agent and how to reach a person. The framework also asks firms to test their communications, which for agents means checking that the disclosure is noticed and the route to a person actually works, not only that the wording exists.
How do you meet the framework?
| Framework area | Control | Evidence it leaves |
|---|---|---|
| Internal governance | Named owner per agent; AI policy; risk assessment process | Owner register; policy; risk records |
| Staff competence | Training for people who oversee or work with agents | Training records |
| Level of human involvement | Probability-severity assessment per agent capability, setting in, over or out of the loop | Capability risk matrix |
| Data quality and lineage | Agent sources approved and recorded; provenance of each fact | Source register; decision records |
| Bias | Outcome testing across groups for capabilities affecting people | Test results |
| Explainability and traceability | Per-action record of facts, criteria, limits and approver | Decision records |
| Repeatability and tuning | Regression testing on real cases after each change | Test reports |
| Stakeholder communication | Disclosure of agent use; review route to a person | Disclosure text; review logs |
Comparable frameworks ask similar questions in different words. The NIST AI Risk Management Framework organises them into four functions, and financial firms in the EU address operational risk under the DORA ICT risk management framework for AI agents.
What will a regulator, auditor or customer ask?
- Who owns each agent, and how are decisions about it made?
- How did you decide the level of human involvement for each capability?
- Can you trace a past action to the facts and limits behind it?
- How do you test the agent after a change?
- What do people affected by the agent know about it, and how do they seek review?
What does the framework not cover?
It is not a legal requirement and creates no penalties. It does not replace the PDPA, MAS rules or sector law. It does not prescribe technical architecture, specific tests or thresholds. For how these sit with binding rules, 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 the framework only for Singapore organisations?
No. It is written for any organisation and is often used as a reference outside Singapore. For organisations in Singapore it signals what regulators consider good practice.
How does human over the loop work for an agent?
A person monitors the agent's activity, receives alerts on unusual outcomes and can intervene or stop it. It suits capabilities with moderate severity and good evidence of reliable performance.
Does adopting the framework satisfy MAS FEAT?
They overlap but are not the same. FEAT is aimed at financial institutions and is more specific on fairness and accountability. Firms supervised by MAS should map their controls to both.