Points of Focus

• MAS published SAFR, a runtime governance framework for autonomous AI agents in finance.
• SAFR routes each agent action to approved, rejected, human review, or flagged.
• Circle contributed a use case alongside Mastercard, Visa, and four others.

The Monetary Authority of Singapore (MAS) published a governance framework on July 3 for autonomous AI agents in financial services, addressing a gap that regulators have so far treated as a future problem.

The white paper, titled “Safeguards for Agentic Finance at Runtime” (SAFR), was developed under MAS’ BuildFin.ai initiative alongside leading financial institutions and fintechs. 

Its scope covers AI agents capable of initiating payments, submitting trading orders, approving credit applications, filing regulatory reports, and settling insurance claims without waiting for human sign-off.

How SAFR works

MAS frames the governance problem precisely: “As AI agents in financial services increasingly carry out tasks autonomously and at speed beyond practical human intervention, financial institutions need real-time safeguards to ensure that the behavior of AI agents remains within predefined mandates, policies and risk boundaries.” 

Existing governance processes, built for human decision-making, do not scale to autonomous systems that execute faster than a compliance officer can read an alert.

SAFR introduces a four-component architecture:

• Agent identity establishes which agent is acting and under what authority. 
• A controls repository holds the predefined rules and risk limits. 
• A disposition engine evaluates each proposed action in real time, routing it to one of four outcomes: approved for automatic execution, rejected outright, sent for human review, or allowed to proceed while flagged for monitoring. 
• An audit log generates a tamper-evident record of every governance decision. 

The result is not a framework where agents ask permission for every transaction. It is one where they check each action against pre-approved rules and escalate to humans only when outside their mandate.

Who built it and what they tested

The paper includes studies from Mastercard, Ant International, Visa, Circle, OCBC and Bank of Singapore, and Manulife across three domains: payments and treasury operations, wealth management and advisory workflows, and client engagement. 

Circle’s presence matters beyond the headline. The company that issues USDC, the dominant MiCA-compliant stablecoin and the foundational payment rail for Circle’s Agent Stack framework published in May 2026, contributed a use case to a central bank’s AI governance framework for autonomous financial transactions. 

SAFR builds on MAS’ Project MindForge AI Risk Management toolkit, extending its model-level risk assessment into a runtime governance layer that operates at the moment of action.

SAFR is voluntary, but banks don’t have to wait

SAFR is an industry reference model, not regulatory guidance. 

No compliance deadline has been set, and institutions can adapt it immediately to their own technology, risk, and compliance stacks without waiting for legislative implementation. MAS said the newly established Future of Finance Institute will support adoption through industry pilots and sandbox experimentation.

The paper addresses the governance of AI agents in financial services at a level of operational specificity no existing regulatory framework, including the EU AI Act or US executive orders on AI, has matched for this use case.