This roundtable, organised by the NUS AI Institute and hosted by IBM Singapore, focused on the responsible deployment of agentic AI in enterprise settings. Agentic systems are able not only to generate outputs in response to prompts, but also to plan, take multi-step actions, invoke tools, interact with other systems and act with varying degrees of autonomy on behalf of users or organisations. This creates opportunities for productivity, service delivery, workflow automation, customer engagement and organisational transformation. It also raises questions of accountability, oversight, liability, technical control, workforce impact and trust, particularly when systems connect to enterprise infrastructure or act across organisational workflows. The session brought together participants from industry, government, academia, civil society, law and finance.

The discussion was framed around a central question: how can organisations in Singapore move from experimentation with agentic AI towards responsible deployment? Participants used three interconnected layers of governance as a starting point for the discussion: principles and ethics provide a strategic “north star”; operational governance assigns responsibility and translates policy into everyday practice; and technical governance supplies the safeguards, monitoring, access controls, testing procedures and runtime controls required once systems are deployed. A recurring premise of the discussion was that principles remain essential, but acquire practical force only when translated into organisational and technical arrangements. As one participant put it: “Rules are fine, but how do you know that the rules are actually being implemented by your AI systems? Especially your Agentic AI systems; especially when they’re spread all across your organisation?” Singapore has not adopted a horizontal AI law, relying instead on targeted interventions in particular areas of harm, practical guidance, industry engagement and co-developed governance frameworks.

For agentic AI, this posture should be translated into the operational mechanisms set out below, with responsibilities and controls made explicit. Governance can then function as enabling infrastructure for confident adoption rather than merely as a brake on it. “Agentic AI” should not, however, be treated as a uniform category. Risk varies with the use case, degree of autonomy, access to data and tools, capacity to transact, reversibility and potential impact. A low-risk internal productivity agent does not warrant the same controls as a system that initiates financial transactions, influences hiring, manages critical infrastructure or interacts with public-facing platforms.

Following the Chatham House Rule, this report synthesises the issues participants regarded as most urgent. It does not attempt to provide a comprehensive legal or technical framework. It first explains the shift from governing systems that produce outputs to governing systems that act, then examines the resulting governance gaps, systemic risks, and recommendations for the different stakeholders.