Power Digital

// AI News

What the MAS AI agent remarks mean for Singapore businesses

Alex Ng
AI News
Illustration with the headline: Mas On Ai Agents And Money

In short: At the Singapore FinTech Festival on 19 November 2026, the Monetary Authority of Singapore (MAS) said AI agents that handle money need bounded autonomy, a clear identity and authority, and specific mandates. If your business deploys agents that touch payments or customer accounts, treat those as your design checklist now.

What happened

MAS Managing Director Chia Der Jiun gave a keynote at the Singapore FinTech Festival 2026 on 19 November. A large part of it was about AI agents: software that acts on a person's or company's behalf, including buying things and moving money.

He said that as agents move into activities with financial consequences, "authority, accountability and control become essential". MAS also advised against giving personal AI agents the login credentials to your financial accounts, because there is not yet enough assurance that this is safe.

He named three problems for a bank or payment provider facing an agent:

  • It may not know whether the agent acts with the customer's full authorisation.
  • It may not know whether the agent stays within its mandate and instructions.
  • It is unclear who is responsible if the agent transacts incorrectly or beyond its instructions.

Key details

SAFR and "know your agent"

The speech drew on SAFR (Safeguards for Agentic Finance at Runtime), and set out these lessons:

  • Bounded autonomy: enough discretion to be useful, inside defined limits.
  • Identity and authority: who the agent is and what it may do must be clearly established.
  • Clear mandates: permitted actions, limits and conditions must be specific.
  • Proportionate controls: the more discretion and impact an agent has, the stronger the assurance and controls.

MAS also described a planned "know your agent" protocol covering who the agent is and whom it represents, what authority it has, what mandate it works under, and what assurance exists about its provenance and reliability. The speech says this would let institutions set terms and scope for agents and create dedicated secure access channels.

Agentic Money Sandbox

With the Future of Finance Institute, MAS is setting up a sandbox that pairs AI agents with programmable money, using two cases:

  • Corporate treasury: a company's agent handles corporate treasury management. The company sets the mandate and the participating banks set the controls.
  • Payments: a customer's agent sends payment instructions. The bank or payment provider verifies the agent's identity and authority before executing them.

MAS announced on 3 November 2026 that the Future of Finance Institute had commenced operations, starting with AI in finance. Its media release lists three pillars: PathFin.ai (knowledge sharing), BuildFin.ai (guidance, good practices, industry standards and common tools, with a focus on the safe use of AI agents) and SafeFin.ai (helping institutions implement MAS' Guidelines on AI Risk Management, including for agentic AI, and enhancing the MindForge AI Risk Management Toolkit). Separately, on 18 November Deputy Prime Minister Gan Kim Yong announced three National AI Missions for financial services, one of them on safe agentic transactions.

What MAS signals

Nothing here is a new law. The speech and release describe lessons, a sandbox and guidance under development. If you are a regulated financial institution, check the Guidelines on AI Risk Management directly. MAS signals the direction: banks will want to know which agent is calling, for whom, and within what limits. A business that builds those answers in early will find it easier to connect to a bank or payment provider later.

What it means for businesses

Most Singapore businesses are not banks, so these remarks do not bind them directly. They matter because your bank, payment provider and customers will apply the same thinking to you. Below is a checklist for a business deploying agents that touch payments or customer accounts. It is Power Digital's reasoning from the lessons and problems in the speech, not an MAS requirement.

MAS pointQuestion to answerWhat to build or decide
Bounded autonomyWhat is the most the agent can do without a human?Hard limits per transaction and per day, set in code and not only in the prompt. Anything above the limit goes to a person.
IdentityCan a third party tell this agent apart from others and from a human?A separate service account for each agent, never a staff member's login. Log which agent did what.
AuthorityWho authorised it, and can you prove it?A recorded consent from the customer or company, with date, scope and how to revoke it.
Specific mandateIs the allowed action list written down?A short mandate document: allowed actions, payees, amounts, conditions. Deny everything else by default.
Proportionate controlsDoes control strength match the damage a mistake could do?Read-only access for lookups, approval steps for payments, extra checks for new payees or changed bank details.
AccountabilityWho is responsible if it goes wrong?A named owner inside your company, a written incident process, and contract wording with vendors on who bears what.
No shared credentialsDoes the agent hold anyone's banking password?Use scoped tokens or API access issued for the agent. Do not store customer banking logins.

Extra checks before go-live

  • Test what happens when a customer, a document or a web page tries to give the agent new instructions. Our article on prompt injection for business owners explains why this matters for agents that move money.
  • Keep an audit trail that a non-technical reviewer can read: request, decision, action, result.
  • Add a kill switch that one person can use without a developer.
  • Decide what the agent does when it is unsure: stop and ask, never guess.
  • Check where customer data goes. Sending personal data to an overseas model provider raises separate PDPA questions.

Our take

This is Power Digital's view. The lessons MAS lists are good engineering practice whether or not a regulator asks for them, and they cost far less to build in at the start than to add after an incident. The "know your agent" idea suggests that agent identity and mandates may become something banks ask to see. We would not wait for a formal rule. If you are planning an agent that touches money, write the mandate and the limits first, then build the agent around them. Our AI agent development service works this way, and the IMDA agentic AI framework checklist for SMEs covers the wider governance side. If the agent connects to your own systems, it also helps to have solid web application development underneath, so that limits and logs live in your backend and not in a prompt.

Sources

Tags

#ai agents #payments #singapore #MAS #compliance

// what we build

Explore Our Services

// AI News

Insights & Resources

Read more →
Back to Articles