An organisation introduces an AI agent to handle internal IT requests. It reads support tickets, checks user accounts, retrieves information from approved systems and resolves routine problems. Some requests are completed without an employee reviewing each action.

But what if it changes the wrong permission or discloses confidential information? Who approved its access, and who is responsible? Unlike a chatbot generating text, an agent can take actions affecting real systems.

Singapore’s Model AI Governance Framework for Agentic AI offers organisations a practical way to address these decisions.

What is agentic AI?

Agentic AI refers to AI systems that can plan and carry out multiple steps towards a defined goal, often by using tools or interacting with other systems. Some require approval before acting; others can complete specified tasks independently.

For example, a conventional AI assistant might draft a response to a customer complaint. An AI agent could retrieve the customer’s order, check a delivery system, update the support ticket and send a reply.

The risk changes when an incorrect answer becomes an action that modifies a record, shares data or triggers a transaction.

What is Singapore’s Model AI Governance Framework for Agentic AI?

Singapore’s Infocomm Media Development Authority (IMDA) launched the Model AI Governance Framework for Agentic AI (MGF for Agentic AI) on 22 January 2026. It applies to organisations building agents or using third-party solutions.

Its May 2026 update added industry feedback, case studies and guidance on multi-agent risks, vendors and automation bias.

The framework is voluntary guidance, not a new binding law. It does not replace applicable laws, and it emphasises human accountability. IMDA’s framework overview describes four connected dimensions.

The four dimensions of the framework

1. Assess and bound risks upfront

Not every task is suitable for the same level of autonomy. An agent summarising public information is different from one changing payroll records or approving payments.

Assess data sensitivity, potential consequences, reversibility and available tools before defining the agent’s limits.

For example, an agent may receive read-only access to a database, permission to use only approved APIs, or a requirement to obtain approval before sending external communications. These restrictions belong in the initial design.

2. Make humans meaningfully accountable

Assigning a person to supervise an agent does not automatically create effective oversight. Define who owns the use case, approves deployment and intervenes when necessary.

The framework also highlights automation bias: people may over-trust a system that has performed reliably in the past. A human approval step has little value if the reviewer routinely clicks “approve” without understanding the proposed action.

Reviewers need enough information and authority to reject or stop actions. High-impact decisions may require advance approval; low-risk tasks may warrant periodic review.

3. Implement technical controls and processes

Governance must be reflected in how the agent actually operates. Useful controls include separate agent identities, least-privilege access, approved tools, action limits, testing, logging and monitoring.

Logs should show actions, tool use, approvals and errors, while avoiding unnecessary retention of sensitive information.

Test before deployment and after meaningful changes. A new API or write permission can change the risk. Gradual rollout helps test controls before wider use.

4. Enable end-user responsibility

Employees should understand when they are interacting with an agent, what it can do and where its limits are. Training should cover when to verify results, challenge actions, escalate problems or take back control.

Staff must also retain the skills needed for meaningful supervision.

What does this look like in practice?

The updated framework includes a case study from Dayos, an enterprise automation company, involving an AI agent that handles internal IT tickets.

Dayos separated actions by risk. Low-impact, reversible tasks such as password resets could be automated and periodically audited. Moderate-risk tasks required human approval before the agent acted. High-impact changes, such as certain permission modifications, were excluded from autonomous execution.

The lesson is that autonomy can be assigned per action, not simply per agent. IMDA’s May 2026 update provides this and other examples.

Where does data governance fit?

AI agents rely on data, access permissions and connections to existing systems. Before giving an agent those capabilities, an organisation should know what information is involved, who is accountable for it and whether the proposed use is permitted.

Consider an agent helping a research team organise datasets. Can it access identifiable participant information? Is it allowed to send data to an external service? Can it change original records, or only produce separate working outputs? Who reviews the results?

These are data governance questions. Classification, ownership, quality, privacy and access rules must inform the agent’s configuration, not just policy documents.

An existing data governance framework provides a foundation; agentic governance extends it to autonomy, tool use and traceability.

How should organisations start?

Start with existing technology, security and risk processes.

1. Keep an agent inventory. Record each agent’s purpose, owner, vendor, connected systems, data sources and permissions.
2. Assess the intended actions. Distinguish reading information from modifying, deleting, sending or approving it.
3. Assign decision rights. Decide who authorises deployment and which actions require named human approval.
4. Set and test controls. Restrict access, test unintended behaviour and maintain useful audit records.
5. Review changes and incidents. Reassess permissions, integrations, performance and oversight as the agent evolves.

For procured agents, clarify vendor responsibilities, customer configuration, data handling and incident support. Procurement does not eliminate organisational accountability.

Final takeaway

Singapore’s framework connects responsible-AI principles with operational decisions.

Capability is not enough. Organisations must define authority, restrict access, maintain oversight and retain evidence.

As AI moves from producing information to taking action, governance must move with it. Autonomy can increase, but accountability must remain clear.

References

• IMDA - Model AI Governance Framework for Agentic AI (official overview)
• IMDA - Launch announcement, 22 January 2026
• IMDA - Updated framework and case studies, 20 May 2026

This article is general guidance, not legal advice. Organisations should review applicable laws and sector-specific requirements when deploying AI agents.