Risks of Autonomous AI Agents Not Yet Fully Understood, Josephine Teo Warns

Six months after Singapore released its governance framework for agentic AI, the minister who oversaw it delivered a sobering reality check: the risks of autonomous agents are still not well understood, and it will take time before the full picture emerges.

Speaking at IBM Think on Tour Singapore on July 21, Minister for Digital Development and Information Josephine Teo told an audience of enterprise leaders that the kinds of errors, risks, and safety issues arising from agentic AI go beyond what the industry is used to. Unlike traditional large language models that respond to prompts, agentic AI can initiate and execute tasks with limited human intervention — opening a new category of failure modes that existing governance tools may not catch.

Trust Bank chief information officer Rajay Rai offered a practical counterpoint: the digital bank is already seeing real returns from AI deployment, including a 50 per cent deflection rate in customer service and minimal complaints. But the optimism came with a caveat. SAF C4 commander and Digital and Intelligence Service chief information officer Guo Jinghua argued that human accountability must remain central, with stricter human oversight required for high-risk use cases. For low-risk tasks like information retrieval, less intervention is needed — but the line between low and high risk is not always clear.

Hans Dekkers, IBM's Asia-Pacific general manager, acknowledged there is no settled answer on who takes ownership when an AI agent makes a mistake. That uncertainty is precisely why Teo called for government-industry collaboration: sharing real-world deployment experience is the only way to understand failure modes before they become systemic problems.

Why it matters for Singapore: Singapore was among the first jurisdictions to publish a governance framework for agentic AI, but frameworks are only as good as their real-world calibration. Teo's admission that the risks are not yet fully understood signals that the current rules are a starting point, not the final word. For Singaporean enterprises deploying AI agents in customer service, logistics, or defence, the message is clear: proceed with caution, keep humans in the loop, and share what breaks. The next phase of AI governance will be written from experience, not theory.