SAFR: A new framework for governing AI agents in financial services

AI agents are moving from recommending financial decisions to executing them. For banks, payment providers and other financial institutions, this creates significant opportunities for efficiency, but also raises important questions about security, accountability and control.

From initiating payments and executing trades to processing insurance claims, autonomous AI systems are beginning to take actions with real financial and regulatory consequences.

The challenge? Traditional governance frameworks were largely designed for human decision-making. Assessments before deployment and audits after transactions occur cannot always prevent an unauthorized or high-risk action from being executed in real time.

This is where Safeguards for Agentic Finance at Runtime (SAFR) comes in.

Developed through the Monetary Authority of Singapore's BuildFin.ai initiative, SAFR introduces an industry reference framework for governing AI agents at the point of action, helping financial institutions verify authority, enforce controls and maintain accountability before an autonomous decision is executed.

Explore the key findings below, or download the full SAFR white paper for a deeper look at the framework and its practical applications.

What is SAFR and how does it work?

SAFR proposes a runtime governance layer that sits between an AI agent and the financial systems it interacts with. Before an agent executes a financial action, its identity, authority and proposed activity are evaluated against defined controls.

The framework brings together four components:

• Agent Identity: Verifies the agent against an identity registry.
• Controls Repository: Defines applicable policies, transaction limits and delegated authority.
• Disposition Engine: Determines whether an action should proceed automatically, be monitored, require human approval or be denied.
• Audit Log: Creates a tamper-evident record of the proposed action and governance decision.

Rather than replacing existing payment systems or compliance infrastructure, SAFR provides a reference approach for enforcing governance before financial actions reach execution.

Why SAFR matters for financial institutions

Financial institutions operate in highly regulated environments where unauthorized transactions, operational failures and insufficient oversight can have significant consequences. As autonomous systems take on greater responsibility, institutions need to maintain control over financial decisions without requiring a human to review every routine action.

The SAFR white paper identifies three important gaps in existing governance approaches.

1. Preventing unauthorized actions before execution

Traditional model risk management largely focuses on evaluating AI systems before deployment, while audits review activity after it occurs.

For financial institutions, this creates a critical gap when autonomous agents can initiate consequential financial actions.

SAFR proposes evaluating each action against defined permissions, policies and transaction limits before execution, helping prevent agents from acting outside their authorized scope.

2. Making human oversight meaningful

One of the white paper's strongest findings concerns human intervention. Existing escalation processes can rely on notifications or alerts without clearly defining who must approve an action, when they need to respond or how the decision is recorded.

SAFR proposes structured escalation with defined reviewer authority, response deadlines and auditable decisions.

For financial institutions, this creates a clearer distinction between routine actions that can proceed autonomously and higher-risk decisions requiring authenticated human approval.

3. Establishing consistent accountability

As institutions introduce AI agents across payments, treasury, compliance and other operations, governance risks becoming fragmented.

SAFR proposes a common structure for evaluating actions and recording decisions, supporting more consistent oversight and accountability across deployments.

SAFR in practice: What the white paper reveals

The white paper includes contributions from Ant International, Circle, HSBC, J.P. Morgan Chase, Manulife, Mastercard, OCBC and Visa, bringing together perspectives from across the financial services ecosystem.

It also examines real-world industry approaches, including Mastercard Agent Pay, Visa Intelligent Commerce and Circle's agent wallets.

These examples explore how organizations can establish agent identities, enforce delegated authority and verify customer intent before an autonomous transaction proceeds.

Visa's approach is particularly interesting: customer instructions are confirmed using a FIDO passkey and translated into rules governing what an AI agent is authorized to do. It highlights how familiar authentication technologies are taking on new roles as financial services move toward agentic transactions.

Where Authsignal fits into the SAFR conversation

A central challenge identified by SAFR is ensuring that human approval is meaningful, authenticated and tied to the action being authorized.

This closely aligns with Authsignal's existing authentication orchestration capabilities. Authsignal supports risk-based step-up authentication, configurable action-level policies and auditable authentication decisions. These capabilities provide a foundation for connecting verified human approval to sensitive actions initiated by AI agents.

The use of passkeys to confirm customer intent also highlights a broader opportunity for authentication to establish trusted authorization beyond login.

Authsignal intends to contribute to the BuildFin.ai working group, focusing on agent identity, authenticated human escalation and cryptographically binding approvals to specific actions.

Building trust into agentic finance

As AI agents gain greater authority within financial services, governance must evolve alongside them. SAFR introduces a practical reference approach for enforcing controls before execution, giving institutions a foundation for balancing autonomy with security, accountability and human oversight.

For financial institutions, the key question is no longer just whether an AI agent can perform an action, but whether it has the authority to do so.

Explore the full SAFR white paper for detailed implementation guidance, industry perspectives and financial services case studies.