The Monetary Authority of Singapore (MAS) expects AI agents to begin moving money for customers and institutions, and wants firms to have controls in place at the moment those agents act. Managing Director Chia Der Jiun set out that view on 9 October 2026 in a keynote address at the INSEAD Digital Finance & Agentic AI Summit, titled "Building Trusted Foundations for Digital Finance".
Most of the address covered tokenised assets, digital money and the infrastructure that connects them. Its final section, on what he called agentic finance, looked further ahead, at software that does more than analyse information; it initiates and executes transactions and other consequential actions for the people and firms it serves. He also said MAS has published Version 1.1 of its Safeguards for Agentic Finance at Runtime (SAFR) white paper, and that the Future of Finance Institute has released an open-source reference implementation of the framework.
The speech does not introduce a new rule. It contains no figures, deadlines or supervisory requirements, and SAFR is an industry white paper, not regulatory guidance.
The Commercial Impact: Frictionless Capital and Spread Erosion
The first implication he raised was commercial. If agents are left to hunt for the best deal on a client's behalf, funds may move between providers more quickly, and for thinner gains. Banks and other firms would then have to compete more aggressively on what they offer and make that case to the client's agent as well as to the client.
"There will also be implications for risk management, particularly in liquidity and collateral," said Chia Der Jiun, Managing Director of the Monetary Authority of Singapore.
The speech did not elaborate on those risks. It did place them in a wider setting: agents operating alongside tokenised assets and digital money, which he said would have significant implications. Earlier, at the same address, he described MAS's work on settlement using tokenised bank deposits and regulated stablecoins through BLOOM, which Visa has joined, and the consultation on legislative amendments to the Single-Currency Stablecoin framework, which would bring stablecoins into the Payment Services Act.
Authority, identity and accountability
The second implication concerned safety. He acknowledged the gains agents could bring in efficiency, access, and new services, but argued that using them safely depends on firm controls over what an agent is allowed to do, how it is identified, and who is accountable for its actions.
"Financial institutions must be able to ensure that AI agents act within the scope of a user's instructions, within institutional policies and risk limits, and in compliance with applicable regulatory requirements," said Chia Der Jiun of the Monetary Authority of Singapore.
That standard places the burden on the institution deploying the agent. It is the same accountability question that arises when AI answers the compliance officer: someone at the firm must own what the system does, and be able to show it.
What SAFR proposes
MAS and a group of financial institutions and FinTechs first published SAFR on 3 July 2026, under the regulator's BuildFin.ai initiative. At the time, MAS said the framework sets governance checkpoints that verify and record an agent's proposed actions before it carries out its tasks, and that it builds on the AI Risk Management toolkit developed under Project Mindforge. Industry members had applied it to agent-assisted payments and treasury operations, wealth management and advisory reviews, and client engagement materials.
In the keynote, SAFR was described as a runtime governance approach with three main elements:
- Know the agent: each agent carries a verified identity and a defined set of permissions.
- Check before acting: a planned action is compared with the client's mandate before it is carried out.
- Leave a trail: the agent's proposed actions are verified and recorded, so they can be audited later.
Version 1.1 includes contributions from a wider group of financial institutions and frontier AI developers, according to the speech, which did not specify what had changed from the July edition. The open-source reference implementation is meant to support experimentation, implementation and further development, and MAS invited research institutions, technology firms and financial institutions to contribute to the next phase.
Where it sits in Singapore's AI rulebook
SAFR sits beside MAS's formal supervisory expectations rather than replacing them. In August, MAS told Parliament that agentic AI was already covered by its proposed Guidelines on AI Risk Management. Two days before the keynote, on 7 October 2026, MAS issued those Guidelines in final form, applying them to all financial institutions and all forms of AI. The Guidelines set what supervisors expect. SAFR offers an industry-built way to apply controls at the point an agent acts.
Near-Term Implementation Roadmap
For banks, payment institutions and wealth managers piloting agents, the three SAFR elements give a practical checklist. Compliance and technology risk teams can confirm that every agent has a registered identity and a defined scope of authority, that its actions are checked against the customer's mandate and the firm's risk limits before execution rather than reviewed afterwards, and that each action leaves an audit trail an examiner could follow.
Treasury and risk functions may test how their liquidity and collateral assumptions would hold if customer balances started moving at agent speed in response to small price differences, the scenario the keynote flagged. SAFR itself is not a regulatory requirement, but the final Guidelines already expect financial institutions to manage AI risks at both the enterprise level and the level of each individual use case.
The technology is still developing rapidly, and many questions remain open; MAS plans to continue working through public-private partnerships to develop standards and best practices. Firms deploying agents should follow the next SAFR iterations closely and consider contributing to them, since the controls that emerge are likely to shape what supervisors look for.
