On 13 October 2026, Taiwan's Financial Supervisory Commission (FSC) issued a consultation paper on the safe use of agentic AI in financial services. It seeks public comment on governance, risk classification and security controls for financial institutions adopting agentic AI. The 30-day consultation is intended to inform discussion and future policy development.
Agentic AI can plan steps, call tools and access data across systems to complete tasks. These capabilities may improve efficiency, but can also create risks such as goals drifting from the original purpose, misuse or overreach of tools, disclosure of sensitive information, and errors cascading through multi-step workflows. The FSC's consultation is based on risk-oriented management, proportionality and technology neutrality.
Three consultation areas
| Area | Direction in the paper | Practical consideration |
|---|---|---|
| Safety governance | Manage each use case as a distinct application scenario. Define its business purpose, responsibilities, data and tool access, permitted actions and exception handling. | Match human review, monitoring, fallback, shutdown, recovery and traceability controls to the risk. |
| Risk classification | Assess data sensitivity, impact of actions, capability boundaries and potential to spread across three dimensions; apply the highest applicable rating. | Carry the rating into approval, pilot, deployment, monitoring and review, and reassess it throughout the lifecycle. |
| Security controls | Treat agentic AI as a non-human identity and workload with the ability to act, and design controls with zero-trust principles in mind. | Cover identity, execution environment, network, application and workload, data, behaviour controls, visibility and analytics. |
Initial checks for financial institutions
When preparing comments or planning deployment, financial institutions and technology providers can start with a specific use case. Identify the agent's identity and accountable owner, data it may read, tools it may call, actions or write operations it may perform, and points for human intervention or emergency shutdown. Then match approval thresholds, permissions, logging, continuous monitoring and recovery exercises to the risk rating. Controls should be testable and connect with existing cybersecurity, outsourcing and internal-control processes.
The FSC says the paper does not replace applicable laws, supervisory rules, outsourcing management or existing internal-control requirements in the financial sector. Because agentic AI technology and use cases continue to evolve, consultation feedback may inform later policy work, guidance and supervisory communication.