The government therefore needs to work with partners to learn from their experience, minister says
[SINGAPORE] The use of autonomous artificial intelligence agents poses some level of risks that are yet to be fully understood, warned Minister for Digital Development and Information Josephine Teo on Tuesday (Jul 21).
Her comments at the IBM Think on Tour Singapore event came six months after the release of the Republic’s governance framework for agentic AI, which seeks to provide guidance to organisations on how to deploy agents responsibly.
“The kinds of errors, the kinds of risks, the kinds of safety issues that arise (from agentic AI) are honestly not completely well understood,” said Teo at the event held at the Sands Expo and Convention Centre.
Unlike traditional large language AI models, agentic AI models are able to initiate and execute tasks with limited human interaction.
Speaking at a panel discussion on the benefits of AI for enterprises, Teo reiterated that the risks created by AI differ from the risks “we are used to”.
“It will be quite a long time before we fully appreciate the range of problems that could be created when agents are unleashed.”
The government therefore needs to work with partners to learn from their experience, allowing it to have a “much better chance of improving our methods”, she said.
In a separate panel discussion on scaling AI across enterprises, Rajay Rai, chief information officer at Trust Bank, noted that the digital bank is reaping the benefits from its deployment of AI.
Since implementing AI for front-facing roles such as customer service through chatbots, Trust Bank is able to get real-time sentiment analysis.
It has also achieved a deflection rate of 50 per cent, while also attaining “single-digit” complaints, he added.
In a call-centre context, a higher deflection rate means that more customers are being diverted to self-service channels.
Accountability for the agents
But as organisations are beginning to reap the benefits of agentic AI through visible returns on investment, they remain wary of the potential risks that entails.
“I think different use cases warrant different levels of interventions and different stringency of governance,” said Guo Jinghua, commander of Singapore Armed Forces C4 and chief information officer at the Digital and Intelligence Service.
For low-risk use cases such as information retrieval and administrative-related tasks, such agents would not require stringent human oversight.
On the other hand, in high-risk use cases where the agents will directly lead to better outcomes and thematic actions, then that is where “man-in-the-loop” will have to be in place, he added.
“But regardless of use cases, there is always human centricity… Human accountability must always be there,” said Guo at a media roundtable.
Hans Dekkers, general manager of IBM Apac, acknowledged that there was no straight answer when it comes to taking ownership of mistakes made by AI agents.
IBM views agent governance across three layers: the data source, the model and the governance surrounding the AI agent, Dekkers added.
Guo noted that AI agents are an “evergreen cycle”, meaning that it is up to the company to be responsible for the agents from development to deployment.
“You cannot (go) hands-off an agent just after it has finished development… responsibilities continue. Accountability continues.”
